INVESTIGATING MULTI-SPECIES INTERACTIONS AND SPATIAL
STRUCTURE OF GUT-BACTERIAL COMMUNITIES USING LIVE IMAGING
by
DEEPIKA SUNDARRAMAN
A DISSERTATION
Presented to the Department of Physics
and the Division of Graduate Studies of the University of Oregon
in partial fulfillment of the requirements
for the degree of
Doctor of Philosophy
September 2022
DISSERTATION APPROVAL PAGE
Student: Deepika Sundarraman
Title: Investigating Multi-Species Interactions and Spatial Structure of Gut-Bacterial
Communities using Live Imaging
This dissertation has been accepted and approved in partial fulfillment of the
requirements for the Doctor of Philosophy degree in the Department of Physics by:
Stephanie A. Majewski Chairperson
Raghuveer Parthasarathy Advisor
Jayson Paulose Core Member
Karen Guillemin Institutional Representative
and
Krista M. Chronister Vice Provost of Graduate Studies
Original approval signatures are on file with the University of Oregon Division of
Graduate Studies.
Degree awarded September 2022
ii
© 2022 Deepika Sundarraman
This work is licensed under a Creative Commons
Attribution-NonCommercial-NoDerivs (United States) License.
iii
DISSERTATION ABSTRACT
Deepika Sundarraman
Doctor of Philosophy
Department of Physics
September 2022
Title: Investigating Multi-Species Interactions and Spatial Structure of Gut-Bacterial
Communities using Live Imaging
Animal intestines harbor hundreds of microbial species that play a crucial
role in host health and development. Despite their importance, many questions
about the rules that govern community assembly in these complex environments
remain unanswered and almost impossible to study in humans. For example, is it
possible to construct multi-species communities from an understanding of pairwise
interactions? What is the role of spatial structure and timing in community
assembly? How do species’ spatial structure and interactions affect the host? We
focus on addressing these questions using a consortium of gut bacterial species,
native to the vertebrate model organism of larval zebrafish. We first characterize
pairwise interactions between a consortium of 5 gut bacterial isolates in 2-species
and 5-species competition experiments using an interaction model and find
evidence for higher order interactions that dampened strong pairwise competition
and enabled coexistence in 5-species communities.
iv
We next focus on a specific pair showing strong pairwise competition in 2-
species experiments. Using light sheet fluorescence microscopy, we illuminate on
the role of spatial structure in the competition between two highly aggregated
species localized in the intestinal midgut, namely strains of genera Aeromonas
(AE) and Enterobacter (EN). We test whether altering aggregation and
localization behavior impact this interaction using a bacterial strain Aeromonas-
MB4, derived from parental AE and composed mostly of planktonic cells that
are anterior localized. When AE-MB4 invades fish colonized with EN, it induces
disaggregation of the highly aggregated EN strain, an effect weakened in the
presence of the other 4 species. Additionally, we observe that AE-MB4 induces
increased inflammation compared to the aggregated parental AE strain, suggesting
possible links between spatial structure and host inflammation.
These studies illustrate the complex ways in which species interact with each
other and impact the host and that multi-species gut bacterial communities are
capable of showing resilience by dampening strong competition effects.
This dissertation contains previously published and unpublished material.
v
CURRICULUM VITAE
NAME OF AUTHOR: Deepika Sundarraman
GRADUATE AND UNDERGRADUATE SCHOOLS ATTENDED:
University of Oregon, Eugene, Oregon
The College of Wooster, Wooster, Ohio
DEGREES AWARDED:
Doctor of Physics, 2022, University of Oregon
Bachelor of Arts, (1) Physics (2) German Studies, 2014
AREAS OF SPECIAL INTEREST:
Biophysics of microbial interactions, 3D live imaging, image analysis,
machine learning
PROFESSIONAL EXPERIENCE:
Graduate Employee, University of Oregon, 2017-22
Teach for India Fellow, Teach for India, 2015-17
GRANTS, AWARDS AND HONORS:
Selected Speaker, Physics of Life Symposium, 2022
First Prize, Oregon Bioengineering Symposium Talk Competition, 2020
Diversity Equity and Inclusion grant, UO Department of Biology, 2018
First Year Fellowship, University of Oregon, 2017
Arthur H. Compton Prize, The College of Wooster, 2014
Copeland Award, The College of Wooster, 2013
Mary A. Sanborne Prize, The College of Wooster, 2013
vi
PUBLICATIONS:
Sundarraman, D., Hay, E. A., Martins, D. M., Shields, D. S., Pettinari,
N. L. & Parthasarathy, R., Higher-Order Interactions Dampen Pairwise
Competition in the Zebrafish Gut Microbiome. mBio, 11: e01667-20 (2020)
Leary, C. C., Lankford, M. & Sundarraman, D., Polarization-based control of
spin-orbit vector modes of light in biphoton interference, Opt. Express 24,
14227-14241 (2016)
Leary, C. C., Lankford, M. & Sundarraman, D., Coupling of spin and orbital
degrees of freedom in tunable Hong-Ou-Mandel interference involving
photons in hybrid spin-orbit modes. Proc. SPIE, Quantum Optics and
Quantum Information Transfer and Processing, 95050T (2015)
Leary, C. C., Gilliss, T. & Sundarraman, D., Bimodal Hong-Ou-Mandel
Interference in Symmetric and Asymmetric Optical Systems, The Rochester
Conferences on Coherence and Quantum Optics and the Quantum
Information and Measurement meeting, Optical Society of America,
M6.05.(2013)
Brückner, U. & Sundarraman, D., Engagement with Civil Society EU India
Relations in the Field of Education and Labour Migration: The Case of
Germany, FPRC Journal, 13 228-237 (2013)
vii
ACKNOWLEDGEMENTS
I appreciate a number of people who have made it possible for me to do
this work. A sincere thanks to my dissertation adviser, Raghu Parthasarathy
who served as an excellent mentor and introduced me to the fascinating world
of biophysics and quantitative image analysis. The consistent guidance and
opportunities for growth and learning have been instrumental in my PhD
experience. All my fellow colleagues, within the Parthasarathy lab and in other
labs affiliated with the META center at UO, energized me with their ideas and
feedback. Working collaboratively with other scientists such as Jarrod Smith and
mentors such as Karen Guillemin has enabled me to push my science to new levels.
I thank my friends, both near and afar, and from different walks of life
for their support and encouragement. I appreciate my family, in particular, my
mother, who encourages me to pursue all my dreams and leads with extraordinary
strength and grace. My sister has always been a source of constant guidance
throughout life.
I’ve been fortunate to have many mentors from different facets of life, both
within and outside of science. I appreciate Cody Leary, Jeremy Wegner and
Rajender Shah for their support at different points in my education.
viii
TABLE OF CONTENTS
Chapter Page
I. INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1. Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2. Model systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3. Quantifying species interactions and types of abundance data . . 4
1.4. Interaction types, models and stability . . . . . . . . . . . . . . . 6
1.5. Spatial structure and its significance in the host gut environment 9
1.6. The host: Mucus and the immune system . . . . . . . . . . . . . 11
1.7. Light sheet fluorescence microscopy . . . . . . . . . . . . . . . . 12
1.8. Image analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
II. MAPPING INTERACTIONS IN MULTI-SPECIES COMMUNITIES . 19
2.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.4. Materials and methods . . . . . . . . . . . . . . . . . . . . . . . 42
III. SPATIAL STRUCTURE AND INTERACTION MECHANISMS . . . . 52
3.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
ix
Chapter Page
3.2. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.3. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.4. Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . 80
IV. IMMUNE RESPONSE FOR DIFFERENT BACTERIAL SPECIES . . 88
4.1. Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.2. TNFα response of bacterial strains. . . . . . . . . . . . . . . . . 90
4.3. Macrophage bacteria interactions . . . . . . . . . . . . . . . . . . 92
4.4. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
V. CONCLUSIONS AND FUTURE WORK . . . . . . . . . . . . . . . . . 96
APPENDIX: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
REFERENCES CITED . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
x
LIST OF FIGURES
Figure Page
1.1. Sample chamber setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.2. Exemplar image showing objects seen in light sheet images . . . . . . . . 15
2.1. The commensal gut bacterial isolates in the 5-species consortium . . . . 21
2.2. Pairwise interactions in 2-species competition . . . . . . . . . . . . . . . 27
2.3. Pairwise interactions in 5-species competition . . . . . . . . . . . . . . . 29
2.4. Diversity in 5-species communities . . . . . . . . . . . . . . . . . . . . . 34
3.1. Mono-association spatial structure of two zebrafish commensal
strains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.2. AE-MB4 affects the spatial structure of an aggregated commensal,
EN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.3. Induced disaggregation of EN when challenged by AE-MB4 . . . . . . . 66
3.4. AE-MB4 competes with the wildtype parental AE strain . . . . . . . . . 69
3.5. Five-species communities promote coexistence and restore spatial
structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.6. Other members dampen AE-MB4 competition with AE . . . . . . . . . 76
4.1. TNFα expression for different strains and combinations . . . . . . . . . 91
4.2. Macrophage swarm activity of AE-MB4 . . . . . . . . . . . . . . . . . . 93
A.1. Bacterial load . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
A.2. Interaction coefficients with uncertainties . . . . . . . . . . . . . . . . . 101
A.3. In vitro interaction coefficients . . . . . . . . . . . . . . . . . . . . . . . 101
A.4. Power law model coefficients for species interactions . . . . . . . . . . . 102
A.5. Linear abundance model interaction coefficients . . . . . . . . . . . . . . 103
xi
Figure Page
A.6. Linear model abundance predictions for 5-species experiments . . . . . . 104
A.7. Aggregation assay for AE and AE-MB4 . . . . . . . . . . . . . . . . . . 105
A.8. Cluster size distributions for AE-MB4, AE and EN . . . . . . . . . . . . 106
A.9. Cluster size histogram for AE, EN and AE-MB4 in
mono-association . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
A.10. Spatial distribution when EN is challenged with AE . . . . . . . . . . . 107
xii
CHAPTER I
INTRODUCTION
1.1. Overview
Microbes are ubiquitous and intimately connected with the environments
they reside in, some of which include soil, oceans, mammalian intestines and even
extreme habitats such as deep sea hydrothermal vents and hot and cold deserts.
They perform a variety of functions in these varied environments, ranging from
biodegradation to nitrogen fixation, carbon dioxide sequestration, plant nutrition
and more [1]. Realization of their significance and functionality has prompted an
increased interest in the role of microbes within host intestines. Recent work has
illustrated that microbes are involved in aspects of host immunity, development,
and a variety of diseases [2–11]. For instance, microbes can be linked to functions
such as dietary fat uptake, proliferation of insulin-producing beta cells responsible
for regulating glucose levels in our body, and development of immune cell types
such as T, B and Treg cells [12–26]. Multiple diseases, including inflammatory
bowel disease(IBD), rheumatoid arthritis and Crohn’s disease have been linked to
dysbiosis of the microbiome [27–31]. As studies continue to uncover the various
ways in which microbes impact our lives, there is still a limited understanding
of the determinants of microbial composition, especially in an environment like
the vertebrate gut, where there is great complexity in chemical composition and
anatomy as well as physical flows [32, 33].
Different parts of the gastro-intestinal tract have been shown to differ in
microbial composition [34], indicative of the various selection pressures at play
1
within the gut. The members of this spatially varying composition comprise a
wide variety of bacteria, archaea, fungi as well as viruses [35]. When cataloging
bacteria alone, almost 2000 distinct taxa have been detected to date [36]. The
enormous diversity and functionality of these microbes warrants the question; how
do these members interact with each other within the vertebrate gut and what
rules govern microbial composition? Several studies have described how the initial
residents can determine which microbes are allowed to cohabit and the potential
benefits to resistance against pathogenic invasion [37, 38]. Controlled studies in
model systems allow us to infer general rules for community structure and gather
information on the mechanisms by which multi-species intestinal communities
assemble and impact the host.
In the era of interdisciplinarity, several different approaches are useful to
tackle this question. As a physicist, the gut with its spatiotemporally varying
features and composition, peristaltic flows and bacterial residents serves as a
fascinating model system to study complex interactions. In this dissertation, I
bring a quantitative perspective to studying inter-species bacterial interactions in
the larval zebrafish gut combining techniques of 3D live imaging, image analysis
and biophysical modeling. In the remaining text of this chapter, I set a foundation
for this work by describing model systems, relevant ecological concepts, interaction
models and tools used in this study.
In the subsequent chapter, Chapter II, adapted from published work,
I characterize inter-species interactions in the larval zebrafish gut using an
interaction model and address whether multi-species interactions can be predicted
from a pairwise additive model.
2
In Chapter III, adapted from work recently submitted for publication, I
delineate a novel mechanism for interaction between two bacterial species that
have been found to have spatially distinct localizations.
In Chapter IV, I zoom out to focus on the host and host-bacteria interactions
by describing some results on immune activity for different gut-bacterial species.
Finally, I draw conclusions from all my PhD work and map out various
directions for further exploration in Chapter V.
1.2. Model systems
Studies in humans have revealed novel insights into how factors such as diet,
social environment and age impact gut microbial composition [39–45]. Although
these studies reveal correlations, experiments in controlled systems are necessary to
find evidence for causation and gain mechanistic insights. Various model organisms
have been used to this end, a few examples being mice, fruit flies (Drosophila
melanogaster), nematodes (Caenorhabditis elegans) and zebrafish (Danio rerio).
Each of these model systems have been exploited for their own advantages.
Mice are often preferred for their genetic similarity to humans, however a single
experiment can often span months with most studies usually timed between the
6-20 week age range. Given the intensive task of maintaining mice for long periods,
some of the largest datasets involve only about 10 mice.
In other model systems such as Drosophila melanogaster and Caenorhabditis
elegans experiments are often more tractable than in mice, since these animals
develop on a shorter time scale i.e approximately days and require less
maintenance. It’s unclear however whether the results in these invertebrate
animals will translate into humans as the intestines of these animals differ more
3
compared to the vertebrate gut [46]. For example, zebrafish possess organs like
the liver and pancreas, which have similar functionality to that in humans [46].
Such specialized organs are missing in Drosophila melanogaster or Caenorhabditis
elegans. For both these organisms, data is obtained from dissection of the gut,
fixation of fish or chemical information from stool samples typically in the form of
RNA or DNA sequences from which it is not possible to infer spatial structure or
dynamic information on microbial communities in the gut.
In recent decades, the zebrafish has turned out to be an excellent model
system to investigate host-microbe interactions within the vertebrate gut. These
organisms are amenable to gnotobiotic techniques i.e. can be easily raised devoid
of microbes, which allows for controlled studies on bacterial interactions. In
addition, its optical transparency at larval stages and a wide variety of transgenic
strains makes it an ideal system to visualize bacterial dynamics and relationships
to different host characteristics [47–49]. Recent work generating fluorescently
tagged bacterial isolates allows for visualization of bacterial dynamics in the
native gut environment in live fish [50]. This work has yielded 3D live imaging
studies illuminating specific spatiotemporal features of gut bacterial populations
and competition dynamics, mechanisms of Vibrio cholera invasion, and effects of
antibiotics on the gut microbiota [51–53]. All these features make larval zebrafish
an ideal system for controlled studies of bacterial interactions and their dynamics.
1.3. Quantifying species interactions and types of abundance data
The German zoologist Ernst Haeckel introduced the term ecology as
the study of the relationships of organisms with their environment. Darwin
pioneered some of the earliest comparative ecological studies in the field that led
4
to revolutionary findings on evolution [54]. During the 20th century, experiments
in ecology involved cataloging species in their natural settings [55–57]. It wasn’t
until the notable work of Robert Paine that simple perturbative experiments on
ecosystems to infer species interactions become popular. Rob Paine’s seminal
experimental studies on tidepool ecosystems paved the way for the idea of a
‘keystone species’ [58–61]. To extract the role of a single species on the ecosystem,
Paine removed all the starfish inhabiting one site and left them undisturbed in
another. He observed that the site that lacked the starfish population saw a loss
of species diversity over a period of two years. As starfish preyed on mussels,
when starfish were present, the mussel population was in check. In the absence
of starfish, the mussel population takes over all the available space, eliminating
other species from the community. In this case, the keystone species starfish
was responsible for maintaining diversity in tidepools. Paine’s species-removal
approach was able to isolate individual species’ effects on the remaining species
in the ecosystem.
In the context of the gut microbiome, a variety of approaches have been
used to infer species effects or interactions. Developments in DNA sequencing
technologies have resulted in easy procurement of relative abundance data, often
derived from stool samples. Relative abundance data is a measure of what fraction
of the total population is composed of a given species, not giving information
about absolute abundance. Such data is used to calculate correlation coefficients,
such as Pearson or Spearman correlations, that indicate whether two species
are likely to be found simultaneously or not. It is often assumed that positively
correlated species have cooperative interactions and negatively correlated species
have competitive interactions, which may not necessarily be accurate [62]. Positive
5
correlations between species can also come from scenarios in which two species
are non-interacting or competing, but show an overall increase in abundance.
Moreover, species interactions are in general asymmetric, i.e. the effect of a species
A on B is not the same or equal to that of B on A. These correlation coefficients
are unable to capture this as they are inherently symmetric.
Compared to relative abundance, absolute abundance data give a thorough
representation of the constituent members and their numbers. Studies collecting
absolute abundance information are, however, difficult, and usually not feasible for
large scale microbiome studies, especially related to humans. Obtaining such data
is however more tractable in in vitro settings or in model system experiments that
sample the entire population of the gut to map interactions.
1.4. Interaction types, models and stability
As described earlier, ecological systems such as the gut microbiome, more
often than not, comprise a large number of players. Among any two species,
interactions may be competitive, cooperative, or neutral. Within this realm of
interaction types, one can also expect asymmetric interaction states such as
amensalism and commensalism, where one species is unaffected while the other
is positively or negatively affected . The simplest interspecies interaction to
characterize is that between a pair of species in isolation. The addition of another
member results in a trio that quickly complicates the landscape for interaction
types. Now, for a trio A,B and C, the interaction between A and B is context
dependent and may be impacted by the presence of the species C. This is true
for each of the different permutations in the trio. As communities become more
diverse, mapping interactions becomes an increasingly complex problem.
6
To measure interactions in such multi-species communities, an approach is
to use a well-defined model with limited assumptions. Since two-species/pairwise
interactions form the foundation for all other interaction types, we can investigate
the question of whether interactions in complex communities are a cumulative
effect of multiple pairwise interactions between species pairs. If such an approach
applies, it simplifies the complexity in the community and allows for a bottom-up
approach to designing synthetic communities by simply using information from
pairwise studies. Interactions that do not stem from a pairwise additive model are
termed higher order or indirect interactions [63].
Investigating the pairwise additivity of interactions has taken precedence
in multiple ecological studies [64–67]. Of particular relevance are those centered
around the gut microbiome, done in various animal model systems. In a study
using C. Elegans, the presence of additional species did not imply higher order
interactions and microbial composition was found to be predictable from
two-species competitions [68]. In a different study involving the Drosophila
microbiome, higher order interactions were found to be present and responsible
for shaping host-fitness traits [69].
Higher order interactions have come up in numerous studies and have been
attributed to stability in complex ecosystems [70–73] The concept of stability
itself has significant bearing in studies of the gut microbiome. Stability has
direct implications for pathogen resistance and a stable resilient microbiome has
been attributed to multiple benefits to host health and well-being [74]. Studies
characterizing species interactions are an essential step to develop a better
understanding of what properties constitute microbiome stability.
7
In recent decades, an increasing number of studies related to modeling
species interactions and community stability have been piloted by physicists.
Theoretical work in Pankaj Mehta’s lab has developed a wide variety of statistical
tools and models to study interaction networks [62, 75, 76]. Some of these tools
have been used to infer interactions and ‘keystone species’ in large microbiome
datasets [62]. Other notable work is the quantitative modeling work by Ned
Wingreen’s group to study metabolic interaction dynamics of microbial consortia
including trade-offs and spatial structure to predict diversity in complex
microbiomes [73, 77, 78]. This work showed that multi-species coexistence is more
widespread than previously known and stems from the many ways in which species
partition metabolic tradeoffs. Experimental work on microbial communities of C.
Elegans and soil ecosystems from Jeff Gore’s lab enabled studies of higher order
interactions in controlled systems, realizing the construction of a desired microbial
consortia [65, 68]. Many of the tools used by physicists in these fields translate
from interaction studies in other areas of physics.
In Chapter II, we use similar quantitative approaches in the form of an
interaction model and investigate the rules governing species interactions in diverse
bacterial communities in a controlled set of species in the vertebrate model system
of larval zebrafish, something not been attempted previously. The study by Lopez
et al. using the invertebrate C. elegans model system found that interactions in
multi-species communities were pairwise additive [68]. These results are in fact
contradictory to those I describe in Chapter II.
8
1.5. Spatial structure and its significance in the host gut environment
Spatial information is thought to be a key determinant of interaction
outcomes. For example, a study comprising three-soil derived bacteria showed that
by simply manipulating spatial structure, competition outcomes were altered [79].
By controlling the microscale spatial structure of three species using a microfluidic
device, the study showed that when the habitats of the species were connected, one
ensured the survival of all species, as opposed to the outcome when their habitats
were isolated [79]. It is also hypothesized that such structuring could explain
observations such as the existence of 4000-500,000 distinct microbial taxa found
in as little as 30g of soil [55, 80, 81]. Spatial structure can thus be responsible for
promoting diversity in complex communities. This may be of particular relevance
in the gut microbiome, where it is known that distinct regions of the intestine
harbor distinct environments [82–85] and that gradients of nutrients and oxygen
can result in microbial reservoirs in regions such as the colonic crypts [84]. How
the hundreds of species in the human intestine exploit these habitats to coexist
and sustain diverse communities remains to be understood. The exploration of
microbial spatial structure within the gut can thus shed insight into mechanisms of
interactions.
Apart from the variations in the locations and the potential composition
of microbes in different regions in the gut, there are also differences relating
to how microbes may socially organize themselves. Members of a species may
be planktonic, i.e. existing as discrete individuals, either motile or non-motile.
Microbes of the same species may also choose to aggregate i.e. form a collective
clump of cells. When such aggregates of cells unify their resources and act as a
collective entity often adhering to surfaces, these clusters of cells are referred to as
9
biofilms. One well known example is the biofilms of the pathogen Pseudomonas
aeruginosa that are often found to be resistant to antibiotics due in part to
this biofilm forming property. [86, 87]. In other cases, bacterial motility can be
beneficial for infection spread and can cause inflammation in cases such as Vibrio
cholerae, Helicobacter pylori and Salmonella typhirium [88].
The spatial organization of bacteria plays a role in determining community
structure as well as their impact on the host, yet its characterization remains
challenging. A vast amount of data in microbiome research uses sequencing
methods applied to fecal samples that don’t retain spatial information. Imaging
of in vivo spatial structure in organisms such as mice entails intestinal dissection
which does not reveal dynamics and assumes that the structure within the gut in
live animals is preserved upon removal. Zebrafish, due to its optical transparency
at larval stages, has emerged as a powerful model to study spatial structure. In our
lab, the live imaging of commensal microbial communities in germ-free zebrafish
has led to the discovery of characteristic spatial domains for certain microbial
members [51–53, 89–91]. Such work has led to the discovery of correlations
between aggregation behaviors and location in the intestine and dynamics that
shape the structure of intestinal communities [91, 92].
The variance in spatial structure across species leads to several unexplored
questions in the sphere of community structure in the gut microbiome. For
instance, what is the connection between spatial structure and competition? How
are species interactions impacted by their spatial distribution? And alternatively,
how is spatial structure impacted by interactions? In Chapter III, adapted
from work currently under revision for publication, I delve into this question by
dissecting the dynamics, spatial structure and interactions between two zebrafish-
10
native gut bacterial species with overlapping spatial domains, Aeromonas veronii
(ZOR0001) (AE) and Enterobacter (ZOR0014) (EN) and two species with distinct
aggregation and distribution, Aermonas-MB4 (AE-MB4) and EN.
1.6. The host: Mucus and the immune system
As previously noted, varied chemical and nutrient gradients exist along the
gut that may be important to determining microbial spatial organization and
interactions. A critical constituent of this environment is host mucus, composed
of glycoproteins and serving multiple functional roles [93]. The lumenal mucosal
layer within the gut acts as a barrier against microbes as well as a source of
nutrition for certain intestinal species [84]. This complex relationship has been
illustrated through several studies. For instance, the pathogen Vibrio cholera
has been found to degrade intestinal mucus [94, 95]. Other bacterial species
are known to stimulate the differentiation of mucus-producing goblet cells [96].
In the case of Pseudomonas aeruginosa, several studies have found that mucin
proteins can stimulate bacterial aggregation and biofilm formation [97, 98]. Thus
mucus-bacteria interactions can have significant consequences for the host and the
microbiota. Through investigations described in Chapter III and IV, I address how
a specific bacterial species (AE-MB4) that is deficient in sensing mucus has altered
interactions compared to its wild-type counterpart (AE), and induces intestinal
inflammation.
The cross-talk between the immune system and the microbiota is another
crucial determinant of host health. Certain microbes have been known to promote
immune system development, while the immune system itself can act as an
ecological filter for microbes [99, 100]. Studies in zebrafish have revealed species-
11
specific immune responses, where a zebrafish commensal Shewenella secretes
potent anti-inflammatory factors that can mute strong inflammation caused by
other abundant pro-inflammatory members [101]. Further, studies also illustrate
that immune responses are linked to motility behaviors of specific strains and
their spatial organization [90]. To explore the connection between the host
immune system, mucus sensing, and bacterial spatial structure, I characterized the
immune response of specific commensal members in single-species and two-species
experiments and qualitatively document unique instances of immune-bacteria
interactions, described in Chapter IV.
1.7. Light sheet fluorescence microscopy
3D imaging at a high resolution is necessary for studies of the dynamics
of bacterial populations and immune cells in the gut of live animals. Some of
the commonly used approaches for 3D imaging are confocal and light sheet
fluorescence microscopy. Confocal imaging works on the principle of exciting the
entire imaging volume using a laser and then scanning a single focal point in
three dimensions while using a pinhole to block any out-of-focus light. Although
this approach can generate images at high resolution, image acquisition is slow
as a single point is being scanned in three dimensions. Advancements to this
approach such as spinning disk confocal can speed up imaging by using multiple
pinholes that scan multiple points simultaneously. Confocal microscopy of all
sorts, however, expose a large volume to large amounts of light which can lead
to phototoxic effects. Such phototoxicity can result in abnormal dynamics of live
organisms as shown in the case of skeletal morphogenesis [102].
12
Our lab uses a light sheet fluorescence microscope to obtain 3D images at a
high resolution for studies of the dynamics of bacterial populations and immune
cells in the gut of live animals. Light sheet fluorescence microscopy involves
illuminating and imaging an entire plane of the sample at a given point of time
and then scanning only in one dimension, perpendicular to the plane, i.e. through
the depth of the specimen. This combination results in high resolution as well
as fast imaging. Since only a single plane is excited at a given time, we reduce
photobleaching and phototoxicity. The microscope in the lab described later has
been customized for imaging bacterial populations in the larval zebrafish gut over
long periods of time. Other light sheet microscopes with varying designs, such as
passing a collimated beam through a cylindrical lens are widely used and described
elsewhere [103–106]. This technique has had a wide variety of applications in
live imaging, with examples ranging from imaging host-microbe interactions to
developmental dynamics and even neuronal firing in live animals [51, 107, 108].
The microscope in the lab comprises a light source (laser) that is scanned
using a galvo mirror oscillating at a frequency of 500 Hz, much faster than the
camera exposure time of 30ms, which results in a time-averaged sheet covering the
entire field of view. The sheet is transmitted through a lens which provides it a
thickness of a few microns, the depth excited within the specimen. Perpendicular
to the sheet, a detection lens captures fluorescence emission. For the imaging
experiments described in this dissertation, lasers at 488nm and 568nm were
used to excite fluorescent proteins in the specimen in series to image pairs of
bacterial species or a combination of immune cells (TNFα+ cells or macrophages)
and bacteria. A single 3D image of an entire fish gut comprising 4 scans in 1
fluorescence channel takes about 45s to obtain. This microscope thus allows
13
one to combine fast image acquisition with high resolution allowing for robust
characterization of bacterial population dynamics in live fish.
FIGURE 1.1. Image showing a live fish mounted within the custom sample
chamber on the lab’s light sheet microscope.
A detailed description of the setup is provided elsewhere [89]. In brief, the
mounting setup comprises a custom designed sample chamber that is filled with
sterile embryo medium and MS-222 anesthetic (20µL/mL). Anesthetized fish are
mounted using agar (0.6% in sterile embryo medium) in glass capillaries as in
Fig 1.1. Each capillary is inserted into the sample chamber and held by a mount
connected to a stage that is controlled with custom-built software. The fish are
extruded from the glass capillary before imaging. For time lapse experiments,
involving imaging fish overnight, up to 6 fish can be mounted for imaging at for
14
up to 16 hours. The data is saved and analyzed using a customized image analysis
pipeline described below.
1.8. Image analysis
The studies described in Ch. III and IV used image analysis approaches
for quantification of spatial structure, dynamics and immune response from
3D images. A single scan through the larval zebrafish gut comprises a set of
2D images of approximate size (2500x1500) pixels, an area of (406x244)µm.
Consecutive 2D images in the scan are taken with a step size of 1 micron. A
typical 3D scan comprises about 150-300 2D images. An exemplar 2D image of
a wild-type zebrafish gut inoculated with EN is shown in Fig 1.2, showing objects
including planktonic or individual bacterial cells, aggregates or clumps of bacterial
cells, autofluorescent zebrafish cells, background autofluorescence that varies due
to the amount of mucus in the gut, and other background noise. The bacteria in
these fish could either be planktonic i.e. exist as single individual cells or form
large multicellular aggregates shown in Fig 1.2.
FIGURE 1.2. An example 2D image (right) of the midgut region from a 3D scan
(left) illustrating some of the different objects detected in these images. The gut is
outlined in yellow. Scale bar: 50µm
15
The general aspects of the image analysis pipeline are described here, while
the specific quantification tasks are described in Ch. II and III.
1.8.1. Individual bacteria detection
Individual bacteria detection is performed in two steps. The first step
involves identifying all potential bacteria objects using a blob detection algorithm.
The second step uses a convolutional neural network to classify all objects as
bacteria and noise blobs. The network was developed and has been described in
detail in earlier published work [109].
The initial process of identifying various potential bacteria in the images
uses a local maxima finder to detect objects of the size of a bacteria within each
2D image of the 3D scan. The 2D masks of identified objects are stitched in the
depth dimension to generate a 3D mask for the image stack. The center of mass of
each of the identified regions is then used to define a (30x30x10) pixel sized voxel
or (5x5x10)µm, a cube large enough to encompass a single bacterial cell. Each
individual voxel from the original image is saved after being normalized to have 0
mean and unit standard deviation before being passed onto the neural network.
Labeling for training the neural network was done using software developed
in Python that displays each saved voxel displaying the potential bacteria object
and allows the user to assign a ‘b’ for bacteria or ‘nb’ for not-bacteria label to
it, thus allowing for binary classification. In prior work, the convolutional neural
network was trained on a bacterial species of the genus Vibrio [109]. For the study
described in Ch. II, a new labeled set of 18,013 images (5239 bacteria and 12774
blobs) was generated for training on Aeromonas-MB4 (AE-MB4) species.
16
The labeled data was input into the network described in [Hay 2019] for
a total of 120 epochs. Cross- validation on the training set was performed to
minimize and assess overfitting to the data. A test set of 1548 bacteria and 11,163
noise blobs yielded a classification accuracy of 91% for AE-MB4.
Each of the models trained is amenable to transfer learning, in which
the network is initially trained on one of these species and then the final fully
connected layer is trained on new data from another species. This approach seems
to work well for training a Vibrio network on Pseudomonas data [109]. For AE-
MB4 however, the accuracy of the classification results was considerably better
when the network was independently trained on the species itself and no transfer
learning was performed. The same network was trained for another species,
Enterobacter (EN) using data labeled specifically for EN. The resulting trained
models for AE-MB4 and EN are applied to the new data for predictions.
The saved x, y, z locations and the predicted label from the network output
are visualized using software written in Python. Objects that are misclassified are
manually filtered during the visualization step.
1.8.2. Gut and aggregate segmentation
For segmentation of the gut and bacterial aggregates in the larval zebrafish
gut in 3D images, we use a widely used algorithm developed for biomedical
images, U-net [110]. Details of the network architecture are provided in [110]. The
implementation for zebrafish gut datasets was done by a previous lab member
which involved slight modifications described in [111]. For the fluorescence
microscopy images, approximately N=1340 2D images of the gut were hand-
labeled. After the initial training, the algorithm is used to identify the mask in
17
a test set and outputs a mask with a binary pixel-wise prediction for each pixel
in the image. A limited amount of post-processing stitches various regions in the
mask. The mask is visualized using custom software and edited if need be.
For images of bacterial aggregates, a combination of thresholding based
on background intensity and manual labeling was done to determine aggregate
objects. This manually labeled data has primarily been used in analysis in Chapter
III.
In the next few chapters, I use tools from image analysis, live imaging and
quantitative modeling to probe (i) the rules governing species interactions in multi-
species communities within the vertebrate gut (Chapter 2), (ii) the connection
between spatial structure and inter-species interactions (Chapter 3) (iii) host
immune response for species with different spatial structure (Chapter 4).
18
CHAPTER II
MAPPING INTERACTIONS IN MULTI-SPECIES COMMUNITIES
This work is adapted from previously published co-authored work by D.
Sundarraman, E. A. Hay, D. M. Martins, D. S. Shields, N. L. Pettinari, and
R. Parthasarathy, Higher-Order Interactions Dampen Pairwise Competition in
the Zebrafish Gut Microbiome, mBio 11, e01667 (2020). I was involved in the
experiments, analysis and writing of this work.
2.1. Introduction
Intestinal microbes exist in complex and heterogeneous communities of
interacting, taxonomically diverse species. The composition of these communities
varies across individuals and is crucial to the health of the host, having been
shown in humans and other animals to be correlated with dietary fat uptake [13,
14], organ development [12, 20], immune regulation [10, 16–18, 21, 23] and a wide
range of diseases [3–9, 11, 15, 22].
Despite the importance of intestinal communities, the determinants of their
composition remain largely unknown. A growing number of studies map the effects
of external perturbations, such as antibiotic drugs [112, 113] and dietary fiber [114]
and fat [19, 115] on the relative abundance of gut microbial species. Intrinsic
inter-microbial interactions, however, are especially challenging to measure and
are important not only for shaping community composition in the absence of
perturbations but also for propagating species-specific perturbations to the rest
of the intestinal ecosystem.
19
The considerable majority of studies of the gut microbiota have been
performed on naturally assembled microbiomes by sequencing DNA extracted from
fecal samples, an approach that provides information about the microbial species
and genes present in the gut, but that imposes several limitations on the inference
of inter-species interactions. The high diversity of natural intestinal communities,
and therefore the low abundance of any given species among the multitude of its
fellow residents, implies that stochastic fluctuations in each species’ abundance will
be large, easily masking true biological interactions. The accuracy of inference
is considerably worse if only relative, rather than absolute, abundance data is
available [62, 116–118], as is typically the case in sequencing-based studies.
Finally, we note that fecal sampling assesses only the microbes that have exited
the host, which may not be representative of the intestinal community [119].
An alternative approach to using DNA sequencing and naturally assembled
host-microbiota systems is to build such systems from the bottom-up using model
organisms. This is accomplished by using techniques for generating initially germ-
free animals, and well-defined sets of small numbers of microbial species, and
then measuring the populations of these species resident in the intestine. Recent
work along these lines has been performed using the nematode Caenorhabditis
elegans [68] and the fruit fly Drosophila melanogaster [69, 120]. However, as
described further below, these studies imply different principles at play in the
different systems. Moreover, it is unclear whether conclusions from either model
platform translate to a vertebrate gut, which has both greater anatomical
complexity and more specific microbial selection [121]. To address this, we measure
bacterial interactions in larval zebrafish (Fig 2.1A), a model vertebrate organism
amenable to gnotobiotic techniques [47–49, 122], which has enabled in earlier work
20
that investigated pairs of bacterial species the discovery of specific interbacterial
competition mechanisms related to intestinal transport [51, 52]. The experiments
described below involve several hundred fish, each with 1-5 resident bacterial
species, enabling robust inference of inter-species interactions.
A
B C
Species
FIGURE 2.1. (A) A 7 day post fertilization larval zebrafish, with a dotted line
outlining the intestine. Scale bar: 500 µm. (B) Chromogenic agar plate showing
colonies of all the candidate five species (AC: milky opaque, AE: reddish purple,
EN: blue, PL: dark purple and PS: colorless translucent) (C) The abundance
per zebrafish gut of each of the five bacterial species when colonized in mono-
association with the host, assessed as colony forming units (CFU) from plated gut
contents. Each circular datapoint is a CFU value from an individual fish (N =
13, 17, 15, 8, and 10, from left to right), with the mean and standard deviation
indicated by the square markers and error bars.
The ability to quantify species abundance and to manipulate it by controlled
addition or subtraction of species is commonplace in macroscopic ecological
21
log10(CFU/gut + 1)
investigations. Its implementation here enables connections between intestinal
microbiome research and a large literature on ecosystem dynamics. An issue
whose importance has been realized for decades is the extent to which interspecies
interactions are pairwise additive, or whether higher-order (often called indirect)
interactions are necessary to explain community structure [123, 124]. The term
’higher-order interactions’ has been defined in various ways in the ecological
literature [124, 125], in some cases referring specifically to non-additive changes in
a species’ growth rate given the presence of additional species or to changes in the
nature of the interaction between two species induced by additional species, but
in others referring more generally to any interaction that cannot be captured by a
pairwise model. We adopt the latter, commonly used, definition, which is agnostic
to underlying mechanisms [63]. In our analyses below, we consider various pairwise
models and assess their ability to describe data from multi-species communities;
mismatch is indicative of the existence of higher-order interactions. Pairwise
additivity, if dominant, simplifies the prediction of ecosystem composition, which
would be desirable for therapeutic applications of microbiome engineering. Higher-
order interactions may stabilize multi-species communities according to several
recent theoretical models described further in the Discussion [70–73], implying that
quantifying and controlling indirect effects may be necessary for reshaping gut
microbiomes.
Whether host-associated or not, microbial communities have shown a
variety of interaction types. A classic study involving cultured protozoan species
found good agreement between the dynamics of four-species consortia and
predictions derived from measurements of pairs of species [64]. Similarly, Friedman
and colleagues showed that the outcomes of competitions among three species
22
communities of soil-derived bacteria could be simply predicted from the outcomes
of pairwise combinations [65]. In contrast, experiments based on the cheese rind
microbiome found significant differences in the genes required for a non-native
E. coli species to persist in a multi-species bacterial community compared to
predictions from pairwise coexistence with community members [67]. A closed
ecosystem consisting of one species each of an algae, bacteria, and ciliate exhibited
a strong non-pairwise interaction, in which the bacteria is abundant in the
presence of each of the algae or ciliate alone, but is subject to strong predation
in the three-species system [66].
Within animals, the interaction types observed in the few studies to date
that make use of controlled microbial communities in gnotobiotic hosts are also
disparate. Competitive outcomes of three-species communities from subsets of
eleven different bacterial species in the gut of the nematode C. elegans could be
predicted from the outcomes of two-species experiments, with indirect effects found
to be weak [68]. In contrast, work using well-defined bacterial assemblies of up to
five species in the fruit fly D. melanogaster found strong higher-order interactions
governing microbe-dependent effects on host traits such as lifespan [69].
To our knowledge, there have been no quantitative assessments of inter-
bacterial interactions using controlled combinations of microbial species in a
vertebrate host, leaving open the question of whether higher-order interactions
are strong, or whether pairwise characterizations suffice to predict intestinal
community structure. We therefore examined larval zebrafish, inoculating initially
germ-free animals with specific subsets of five different species of zebrafish-
derived bacteria and assessing their subsequent absolute abundances. Though the
number of species is considerably fewer than the hundreds that may be present
23
in a normal zebrafish intestine, it is large enough to sample a range of higher-
order interactions, yet small enough that the number of permutations of species
is tractable.
As detailed below, we find strong pairwise interactions between certain
bacterial species. However, we find weaker interactions and a greater than
expected level of coexistence in fish colonized by four or five bacterial species. This
suggests that measurements of pairwise inter-microbial interactions are insufficient
to predict the composition of multi-species gut communities, and that higher-order
interactions may dampen strong competition and facilitate diversity in a vertebrate
intestine.
2.2. Results
Zebrafish (Fig 2.1A) were derived to be germ-free, and then were inoculated
at 5 days post-fertilization (dpf) with the desired combination of microbial
species by addition of bacteria to the flasks housing the fish. Approximately 48
hours later, fish were euthanized and their intestines were removed by dissection.
Intestines and their contents were homogenized, diluted, and plated onto
chromogenic agar (Methods). Secreted enzymes from each of the five candidate
bacterial species generate particular colors due to substrates in the chromogenic
medium, allowing quantification of colony forming units (CFUs) and therefore
absolute intestinal abundance (Fig 2.1B).
The five species examined were selected as diverse representatives of genera
commonly found in the zebrafish intestine. Full names and species identifiers are
given in Methods; we will refer to these through most of the text by genus name or
two letter abbreviation: Acinetobacter (AC), Aeromonas (AE), Enterobacter (EN),
24
Plesiomonas (PL), and Pseudomonas (PS). As expected given their association
with the zebrafish gut microbiome, each species in mono-association, i.e. as the
sole species inoculated in germ-free fish, colonizes robustly to an abundance of
103-104 CFU/gut, corresponding to an in vivo density of approximately 109-1010
bacteria/ml (Fig 2.1C).
Pairwise interactions in di-associations.
We first examined all ten possible co-inoculations of two species, which
enables assessment of pairwise interactions in the absence of higher-order effects.
Intestinal CFU data shows a wide range of outcomes for different species pairs. As
exemplars, the CFUs per gut for each of two species, AC and EN, in the presence
of each of the other four are displayed in Fig 2.2A and Fig 2.2B respectively. The
abundance of AC is similar in the presence of any second species to its value in
mono-association. In contrast, the mean EN abundance is similar to its mono-
association value if co-inoculated with PL or PS, about 10 times lower if co-
inoculated with AC, and over two orders of magnitude lower if co-inoculated with
AE, implying in the latter cases strong negative interactions.
Parameterizing the strength of interactions between species is necessarily
model dependent, contingent on the functional form of the relationship between
one species’ abundance and the other’s. We show that the conclusions we reach
regarding interaction strengths, especially their shifts when multiple species are
present, are qualitatively similar and therefore robust for a wide range of models.
We first consider a phenomenological interaction coefficient CIIij that is linear in
25
log-abundance, characterizing the effect of species j on species i as:
log P II10 i = ⟨log10 P Ii ⟩+ CII IIij log10 Pj (2.1)
where Pi denotes the abundance of species i and the superscript I or II denotes a
mono- or di-association experiment. This form is motivated by the distribution of
gut bacterial abundances being roughly log-normal, with species addition capable
of inducing orders-of-magnitude changes (Fig 2.2A,B). This CIIij can be derived
as the interaction parameter in a competitive Lotka-Volterra model modified to
act on log-abundances (Methods). Qualitatively, a positive Cij implies that the
abundance of species i increases in the presence of j. Similarly a negative Cij
indicates that the abundance of species i declines in the presence of species j.
Subsampling from the measured sets of bacterial abundances gives the mean and
standard deviation of the estimated interaction parameters (Methods).
We plot in Fig 2.2C the CIIij defined by Eq. 3.1 calculated from all di-
association data of all species pairs (N = 190 fish in total). For determining CIIij ,
we only use data from fish in which both species were detected so that abundance
changes of one species can definitively be ascribed to the presence of the other
within the gut. Uncertainties in CIIij are estimated from bootstrap subsampling
(see Methods). The interactions are predominantly negative. Thirteen out of
twenty coefficients differ from zero by over three standard deviations, indicating
both a large magnitude and a less than 0.001 probability that the interaction
strength is zero or of the opposite sign. The total bacterial load, i.e. the sum of
the bacterial abundances, is similar for all the di-associations suggesting that the
interaction effects do not stem from changes in intestinal capacity (Fig 2.2D).
26
A B
- -
mo
no n ono n
cia
tio m atio
ssoa Second Species sso
ci
a Second Species
C Interaction Coefficients D Bacterial Load
Species j Species j
FIGURE 2.2. Abundances per zebrafish gut of (A) AC and (B) EN in mono-
association (grey) and in di-association with each of the other bacterial species
(blue/green). Each circular datapoint is a CFU value from an individual fish ((A)
N = 13, 21, 19, 20, and 27 and (B) N = 15, 19, 22, 18, and 23 from left to right),
with the mean and standard deviation indicated by the square markers and error
bars. (C) Matrix of pairwise interaction coefficients CIIij characterizing the effect of
species j on the abundance of species i. Coefficients that differ from zero by more
than three standard deviations (provided in Fig A.2A) are outlined in black. (D)
The average bacterial load per zebrafish in each of the di-association combinations,
expressed as log10 of total CFUs. The standard deviations are between 0.3 and 1.1
and are displayed in Fig A.2A). Values on the diagonal are the mono-association
load for each of the five species.
27
Species i AC  log10(CFU/gut + 1)
Species i EN log10(CFU/gut + 1)
Though the physical and chemical environment of the zebrafish gut is
likely very dissimilar to test tubes of standard growth media, we examined
abundances of each of the pairs of species in in vitro competition experiments,
growing overnight cultures in Lysogeny broth (LB) media and plating for CFUs
(see Methods). Assessing CIIij as above, we find, as expected, that interaction
coefficients calculated from the in vitro experiments are markedly different than
those measured in vivo (see Fig A.3 and Fig 2.3B).
Our characterizations of interactions within the zebrafish gut are not
qualitatively altered by using a more general power law model to compute CIIij
from absolute abundance data, discussed below (Interactions under more general
models) following the presentation of measurements of interactions between more
than two species.
2.2.1. Pairwise interactions in multi-species communities.
To assess whether the strong competitive interactions we found in two-species
experiments are conserved in multi-species communities, we quantified pairwise
interactions in experiments inoculating fish with four or five bacterial species.
To assess CVij , we adopted a method similar to the leave-one-out approach often
used in macroscopic ecological studies, dating at least to classic experiments in
which single species were removed from tide pools and the abundances of the
remaining species were measured to evaluate inter-species interactions [58]. Here,
we performed co-inoculation experiments leaving out one of the five species of
bacteria and compared intestinal abundances for these four-species communities
to those measured in five-species co-inoculation experiments.
28
A B
Species absent Species j
D
C
II
Cij
CVij
CIIij α
FIGURE 2.3. (A) Abundance per zebrafish gut of one of the bacterial species,
EN, when all five species are co-inoculated (gray) and in each four species co-
inoculation experiment (green) with the omitted species indicated on the axis.
Each circular datapoint is a CFU value from an individual fish (N = 40, 12, 12,
11, and 9, from left to right), with the mean and standard deviation indicated by
the square marker and error bars. (B) Matrix of pairwise interaction coefficients
CVij when 5 bacterial species are present. The coefficients outlined in black differ
from zero by over three standard deviations (see Fig A.2B). (C) The pairwise
interaction coefficients inferred from 4-5 species experiments versus those from 1-2
species experiments. The colors label species i for each interaction pair. (D) The
minimum interaction coefficient calculated from a power-law interaction model for
different values of the exponent α for the 1-2 species (square filled markers) and
the 4-5 species (square markers) experiments.
29
EN log10(CFU/gut + 1)
CVij 
log10 [-min{Cij}] Species i
In approximately N = 10 fish each, we performed all five different co-
inoculations of four bacterial species. The difference in the abundance of species
i in fish inoculated with all five species compared to fish inoculated with four,
missing species j, gives a measure of the impact of species j on species i in the
multi-species environment. As an example, EN abundance in inoculations lacking
AC, AE, PL, and PS, and in five-species inoculations, are shown in Fig 2.3A.
In contrast to the di-association experiments (Fig 2.2B), we see that EN does
not show large abundance differences, in either its mean or its distribution, as a
result of any fifth species being present. Independent of any model, this suggests
that non-pairwise, i.e. higher-order, interactions are present in the multi-species
community.
Again, a variety of options are possible for quantifying interaction coefficients
in the multi-species system. We first consider interaction coefficients as modifying
mean-log-abundances, analogous to the pairwise model of Eq. 3.1:
log V10 Pi = log10 P
IV + CVi ij log10 P
V
j (2.2)
The interaction coefficients CVij that we obtain, displayed in Fig 2.3B, are
different and in general considerably weaker than the CIIij found in the two-species
case Fig 2.2C. There are only three interactions that differ from zero by over three
standard deviations. Strikingly, all three of these interactions are positive. This
shift towards weaker and more positive interactions between the two-species and
multi-species interactions is further illustrated in Fig 2.3C in which the multi-
species interaction coefficients, CVij are plotted against the 2-species interaction
coefficients, CIIij .
30
2.2.2. Interactions under more general models.
As noted, a model that is linearly additive in logarithmic abundances is only
one of an infinite number of choices, and moreover may not adequately capture
the complexity of interactions in the gut. Earlier experiments investigating the
spatial structure of specific microbial communities in the larval zebrafish intestine
have shown that species such as AE, EN and PS form dense three-dimensional
aggregates [91]. The size and location of aggregates and the locations of cells,
conspecific or otherwise, within these aggregates may impact their interactions in
ways that could be sub-linear, linear, or super-linear in population size. Previous
work has also established that gut bacteria may also influence intestinal mechanics
[52], highlighting one of many possible indirect interaction mechanisms whose
functional forms are unknown. Furthermore, other studies have shown that
different modes of physical and chemical communication could result in long range
interactions between different species [126–128]. To address these possibilities, we
evaluated species interactions with a more general power law model, wherein the
interaction effects between species could be non-linear in the abundance of the
effector species. Here the interaction coefficient Cij depends on a power, α, of the
abundance of the effector species j, which we evaluate in the range α = 0.1 to
2, spanning sub-linear and super-linear interactions. Modified versions of Eq. 3.1
and 3.2 give:
P II = ⟨P I⟩+ CII(P II αi i ij j ) (2.3)
and
P V = ⟨P IV ⟩+ CV (P V αi i ij j ) (2.4)
31
from which we can evaluate CIIij and C
V
ij respectively. Note that α = 1 in Eq. 3.3,
2.4, i.e. interactions that are linear in abundance, is simply the steady-state
behavior of the competitive Lotka Volterra model commonly used in population
modeling and are shown in Fig A.5. We provide the CIIand CVij ij for several
different α in Fig A.4. Throughout, as in the logarithmic model shown above,
pairwise interactions in di-association are in many cases strongly negative, while
the multi-species interactions are weaker. This is summarized by studying the
trends in the most negative CIIij and C
V
ij for different values of α, depicted in Fig
2.3D, which shows that for all α the strongest CVij is significantly weaker than the
strongest CIIij , suggesting that our results are robust to choice of model.
2.2.3. Five-species coexistence.
We next consider co-inoculation of all five bacterial species. Examination
of over 200 fish shows a large variety in abundances, depicted in Fig 2.4A as the
relative abundance of each species in each larval gut. Multiple species are able to
coexist, with the median number of species present being 4 (Fig 2.4B). The mean
total bacterial load as well as its distribution (Fig 2.4C) is similar to the mean
and distribution of the mono- and di-association experiments, as well as four-
species co-inoculation experiments discussed earlier. We calculated the expected
abundance of each bacterial species, if the interactions governing the five-species
community were simply a linear combination of the pairwise interactions governing
di-associations, CIIij . Any of the additive models we evaluated can be extended to
combinations of species. Considering the model focused on above, with interaction
coefficients modifying log abundances, the predicted abundance of species i in the
presence of another species j is given by
32
∑
log10 P
V
i = ⟨log10 P Ii ⟩+ CII Vij log10 Pj (2.5)
j ̸=i
where the superscript V denotes the five-species co-inoculation experiment. A
model linear in species abundance (α = 1) is also considered in Methods), and
gives qualitatively similar outputs and conclusions. Sampling from the measured
distributions of each of the interaction coefficients and the mean abundance in
mono-association allows calculation of the distribution of expected P Vi values.
We plot the measured and predicted distributions of intestinal abundances
of each of the five species for the five-species co-inoculation experiment in Fig
2.4D. The measured distributions of each of the species are very similar to each
other. In contrast, the distributions of the predicted abundances vary significantly
by species. For two of the species, AC and PS, the mean of the observed and
predicted distributions are similar. For the other three species, in contrast,
the observed and predicted populations are in strong disagreement, with the
pairwise prediction being at least an order of magnitude lower than the observed
abundances. For EN and PL in particular, we would expect extinction in a large
fraction of fish due to strong negative pairwise interactions; in actuality, both
species are common and abundant.
Similarly, we can extract from the model the predicted frequency of
occurrence of each of the species, regardless of abundance i.e. the fraction of fish
with a non-zero population of that species. We find that the predicted frequency is
much lower than the experimentally observed frequency for PL and EN (Fig 2.4E).
33
FIGURE 2.4. (A) Stacked bar plot of the relative abundances of the five bacterial
species when all five were co-inoculated. Each bar is from a single dissected fish.
The bars are ordered by total bacterial load. (B) Histogram of the total number
of bacterial species present in the gut when all five species were co-inoculated. (C)
The total bacterial load as a function of the number of inoculated species. Each
circular datapoint is a CFU value from an individual fish (N = 63, 232, 187, and
202, from left to right), with the mean and standard deviation indicated by the
square marker and error bars.(D) The predicted (blue Xs) and measured (brown
circles) abundances of each bacterial species in the zebrafish gut when all five
species are co-inoculated. Predictions are based on an interaction model that is
linear in log-abundance using the pairwise CIIij coefficients, as described in the
main text. Solid square markers indicate the mean and standard deviation of the
distributions excluding null counts. The dotted line indicates the experimental
detection limit of 25 cells. The experimental data is from N = 202 fish in total and
the predicted distributions arise from 250 samples of the distribution of interaction
coefficients. (E) The observed frequency of occurrence in the gut from the five-
species co-inoculation experiment versus the predicted frequencies for each of the
five species. (F) The Pearson correlation coefficients calculated from the relative
abundances of pairs of species when all five species were co-inoculated.
34
A B
Number of species
C D Prediction Data
Number of species inoculated Species
E F
Predicted Frequency Species j
FIGURE 2.4.
35
Relative Abundance
Total Load log10(CFU/gut + 1)
Measured Frequency
Species i Log10(CFU/gut + 1)
Frequency
By measuring absolute abundances of bacterial populations in the gut, we
provide direct assessments of inter-species interactions. More common sequencing-
based methods, applied for example to the human gut microbiome, typically
provide relative measures of species abundance, i.e. each taxonomic unit’s fraction
of the total load. Correlations among relative abundances are often used as
measures of interaction strengths [129, 130]. Calculating the Pearson correlation
coefficients of the relative abundances of each pair of species in fish inoculated with
all five bacterial species, we find a strikingly different interaction matrix (Fig 2.4F)
than that inferred from absolute abundance changes (Fig 2.3B), with many strong
negative values. There are many likely reasons for the difference between Pearson
correlations and our more directly measured interaction coefficients. The Pearson
r necessarily attributes correlations between pairs of species as being indicative of
the dynamics of that pair independent of other species, is confounded by overall
changes in total bacterial load, and, perhaps most importantly, is necessarily
symmetric (Cij = Cji). Our Cij inferred from absolute abundance data are notably
asymmetric (Fig 2.2C,Fig 2.3B).
2.3. Discussion
Using a model system comprising five commensal bacterial species in
the larval zebrafish intestine, we have characterized aspects of gut microbiome
assembly. Controlled combinations of inoculated species and measurements of
absolute abundance in the gut, both challenging to perform in other vertebrate
systems, reveal clear signatures of interactions among species. We find strong,
competitive interactions among certain pairs in fish inoculated with two bacterial
species. In contrast, pairwise interactions are weak in intestines colonized by four
36
to five species, and all species are present at equal or greater abundance than
would be predicted based on two-species data.
Our quantification of interaction strengths relies on a minimal set of
assumptions that serve as a general test of additive models. Interaction strengths
are necessarily parameters of some model. In the text, we make use of a model in
which the log-transformed population of a species is a linear function of the other
species’ log-transformed populations, and a more general power law model that
spans both sub-linear and super-linear dependences on population sizes. There are
good reasons to be skeptical of such frameworks. First, intestinal populations may
not be well described by equilibrium, steady-state values. Second, these models
lack spatial structure information. In vivo microscopy of one or two species in the
zebrafish gut [51–53] underscores both of these concerns: populations are very
dynamic with rapid growth and stochastic expulsions; interactions can be mediated
by complex intestinal mechanics; and aggregation and localization behaviors are
species-specific.
Imaging also, however, provides justifications for these rough models. Prior
microscopy-based studies have shown that growth rates are rapid, with populations
reaching carrying capacities within roughly 12 hours [51, 53], well below the 48
hour assessment time considered here. Because of strong aggregation observed
in nearly all bacterial species, most individual bacteria residing in the bulk of
clusters will not directly interact with other species, leading to interactions that
are sub-linear in population size, suggesting a logarithmic or α < 1 power-law
functional form. Furthermore, stochastic dynamics can be mapped onto robust
average properties for populations [51, 131]. It is therefore reasonable to make use
of simple models, not as rigorous descriptions of the system but as approximations
37
whose parameters characterize effective behaviors. We note that all these issues
also affect more commonly used models, such as standard competitive Lotka-
Volterra models that are linear in population sizes. These models are often applied
to gut microbiome data and used to infer interaction parameters [62, 132, 133]
despite a lack of information about their realism. The power law model of
interactions provides the strongest indication of the generality of our conclusions.
Over a range of interaction forms extending from highly sublinear (α = 0.1) to
super-linear (α = 2.0), strong competitive interactions are damped when four or
five species are present (Fig 2.3D), suggestive of higher-order interactions among
intestinal bacteria.
The ecological potential for higher-order or indirect interactions, i.e.
interactions that cannot be reduced to pairwise additive components but rather
result from the activities of three or more species, to be important determinants
of community structure has been appreciated for decades [64, 123, 124].
Identification of higher-order interactions among constituent species is important
for accurate prediction of responses to ecological perturbations such as species
invasion or extinction, as well as functions of multi-species communities, as such
features will not be adequately forecast by examination of direct interactions in
subsystems [123, 134].
Inferring and quantifying indirect interactions in natural ecosystems is,
however, challenging, calling for subtle and model-dependent statistical tests [123,
124, 135]. Constructed or manipulated systems enable more straightforward
assays in which particular species are introduced or removed amid a backdrop
of others. Several such systems involving macroscopic organisms [136–140], as
well as microorganisms [66, 69] have uncovered significant indirect interactions.
38
However, some studies of microbial communities have found weak or negligible
higher-order interactions [64, 65], including one study examining combinations of
species introduced to the C. elegans gut [68]. The complexity of interactions in a
vertebrate gut has remained unclear, and correlation-based methods for inferring
interactions from sequencing-based data have assumed that pairwise interactions
suffice [129, 141, 142].
Our measurements using gnotobiotic larval zebrafish, a model vertebrate,
show strong pairwise interactions when only two bacterial species are present
in the intestine and weak pairwise interactions when four to five species are
present, indicating that higher-order interactions are important (Fig 2.3) . In many
cases, the effect is evident from the raw data itself. For example, EN is strongly
suppressed by AE if the two are inoculated together (Fig 2.2B). Comparing EN
abundance in fish colonized by all species except AE with its abundance in fish
colonized by all species, however, shows little difference, indicating that the EN-
AE interaction is strongly attenuated by the presence of the other bacterial groups
(Fig 2.3A).
Two additional observations also imply the presence of strong higher-order
interactions in our intestinal ecosystem. Considering fish colonized by all five
bacterial species, the mean abundance of each species is at least as high as would
be predicted solely from direct interactions (Fig 2.4D). Moreover, the diversity of
bacterial species is higher would be predicted, as all of the species occur in more
than 50% of fish, contrary to prediction (Fig 2.4E). Considering fish colonized by
all five bacterial species, the abundance of each species is at least as high and the
diversity of bacterial species is higher than the values that would be predicted
solely from direct interactions.
39
Our finding of increased species diversity than expected from pairwise
interactions in a system of several gut bacterial species is consistent with recent
theoretical studies that suggest, for a variety of reasons, that higher-order
interactions are likely to stabilize communities and promote coexistence. The
topic of multi-species coexistence has a long history in ecology. Especially since
classic work by Robert May showing that a system comprising pairwise interacting
constituents will, in general, be less stable as the number of species increases [143],
explaining how complex communities can exist has been a theoretical challenge.
There are many resolutions to this paradox, such as spatial heterogeneity,
interactions across trophic levels, and temporal variation. However, even without
such additional structure, incorporating higher-order terms into general random
competitive interaction models leads to widespread coexistence [70–72]. Such
large-scale coexistence can also emerge naturally from contemporary resource
competition models [73, 144], in which cross-feeding or metabolic tradeoffs
necessarily involve multiple interacting species. Intriguingly, the abundance
distributions of all five of our gut bacterial species, when inoculated together,
are similar to one another. The average Shannon entropy of the five species
community (H = 1.16 ± 0.24) (see Methods) also resembles that of a purely
neutrally assembled community (H = 1.61), reminiscent of dynamics mimicking
neutral assembly that emerge from multi-species dynamics driven by resource use
constraints [73, 145].
Our findings imply that measurements of two-species interactions among
microbial residents of the vertebrate gut are likely to be insufficient for predictions
of community dynamics and composition. Moreover, they imply that inference
from microbiome data of inter-species interactions, for example by fitting Lotka-
40
Volterra-type models with pairwise interaction terms [62, 132, 133, 146] should
not be thought of as representing fundamental pairwise interactions that would
be manifested, for example, if the constituent species were isolated, but rather as
effective interactions in a complex milieu.
Our measurements do not shed light on what mechanisms give rise to
higher-order interactions in our system. Likely candidates include metabolic
interactions among the species, interactions mediated by host activities such as
immune responses, and modulation of spatial structure by coexisting species.
Immune responses are sensitive to specific bacterial species [101] and to bacterial
behaviors [90]. Regarding spatial structure in particular, in vivo imaging of these
bacterial species in mono-association has shown robust aggregation behaviors
that correlate with location in the gut [91] Given the physical constraints of the
intestinal environment, we think that modification of spatial organization due to
the presence of species with overlapping distributions is a likely mechanism for
higher-order interactions. Notably, both immune responses and spatial structure
are amenable to live imaging in larval zebrafish [51, 52, 91]. Though the parameter
space of transgenic hosts, fluorescent labels, and interaction timescales to explore
in imaging studies is potentially very large, such future studies are likely to
yield valuable insights into the mechanisms orchestrating the strong interactions
observed here. Furthermore, examination of the roles of priority effects and other
aspects of initial colonization, as well as stability of diverse communities with
respect to invasion, may reveal potential routes for intentionally manipulating the
vertebrate microbiome to engineer desired traits.
41
2.4. Materials and methods
2.4.1. Animal Care.
All experiments with zebrafish were done in accordance with protocols
approved by the University of Oregon Institutional Animal Care and Use
Committee and following standard protocols [147].
2.4.2. Gnotobiology.
Wild-type (ABCxTU strain) larval zebrafish (Danio rerio) were derived
germ-free as described in [47]. In brief, embryos were washed at approximately
7 hours post-fertilization with antibiotic, bleach, and iodine solutions and then
moved to tissue culture flasks of 15mL sterile embryo medium solution with
approximately 1mL of sterile solution per larva. The flasks were then stored in a
temperature-controlled room maintained at 28◦C.
2.4.3. Bacterial Strains and Culture.
The five bacterial strains used in this study, namely Aeromonas sp.
(ZOR0001), Pseudomonas mendocina (ZWU0006), Acinetobacter calcoaceticus
(ZOR0008), Enterobacter sp. (ZOR0014), and Plesiomonas sp. (ZOR0011) were
originally isolated from the zebrafish intestine and have been fluorescently labelled
to express GFP and dTomato facilitating their identification in our experimental
assays [50, 148]. Stocks of bacteria were maintained in 25% glycerol at −80◦ C.
42
2.4.4. Inoculation of tissue culture flasks.
One day prior to inoculation of the tissue culture flasks, bacteria from frozen
glycerol stocks were shaken overnight in Lysogeny Broth (LB media, 10 g/L NaCl,
5 g/L yeast extract, 12 g/L tryptone, 1 g/L glucose) and grown for 16 h overnight
at 30◦C. Samples of 1mL of each of the overnight cultures were washed twice by
centrifuging at 7000g/rpm for 2 min, removing the supernatant, and adding 1mL
of fresh sterile embryo media. At 5 dpf, the tissue culture flasks were inoculated
with this solution at a concentration of 106 CFU/mL. For each of the competition
experiments involving 2, 4 and 5 bacterial species, equal concentrations were
inoculated into the flasks. After inoculation, the flasks were maintained at 30◦C
until dissection at 7 dpf.
2.4.5. Dissection and Plating.
To determine the intestinal abundance of bacterial species, dissections
of larval zebrafish were performed at 7 dpf. Zebrafish were euthanized by
hypothermal shock. Intestines were removed by dissection and placed in 500µL
of sterile embryo media and homogenized with zirconium oxide beads using a
bullet blender. The homogenized gut solution was diluted to 10−1 and 10−2, and
100µL of these dilutions were spread onto agar plates. For mono- and di-associated
inoculations, tryptic soy agar (TSA) plates were used in which fluorescence could
be used to differentiate up to two species. For inoculations of more than two
species, Universal HiChrome Agar (Sigma-Aldrich) plates were used, allowing
for visual differentiation of each species using a colorimetric indicator. The
abundances of each of the species in the zebrafish gut was determined by counting
the colony forming units on the plates.
43
2.4.6. In vitro competition experiments.
To determine the in vitro competition coefficients, all the different pairwise
combinations of the five species were grown in overnight cultures of LB media
as above. On the following day, cultures were plated at 10−7 or 10−6 dilutions,
depending on the ability to detect both species in a given dilution. Abundances
were obtained by counting the number of CFUs of each species on the plates.
2.4.7. Interaction Models
2.4.7.1. Pairwise Interactions
As noted in the main text, we aim to characterize the measured abundances
of microbial species with numerical parameters that characterize their interactions.
These parameters may simply be phenomenological, but may also relate to
dynamical population models.
Considering two species, labeled by subscripts i and j, the commonly used
competitive Lotka-Volterra model describes the behavior of the populations P by
the following relationship:
( )
dPi − Pi − CijPj= riPi 1 (2.6)
dt Ki
where ri is the intrinsic growth rate, Ki is the carrying capacity, and Cij, defined
for is the interaction coefficient, negative for competitive interactions and positive
for beneficial interactions.
As steady state, dPi = 0, hence,
dt
Pi −Ki
Cij = . (2.7)
Pj
44
Note that this holds even if the initial factor in the Lotka-Volterra equation,
riPi is replaced by any generic function of Pi . As discussed in the main text,
each of the bacterial species examined is known to reach its carrying capacity
approximately 10-12 hours after inoculation [51], a timescale considerably shorter
than the 48 hours post-inoculation of our abundance measurements. It is therefore
reasonable to make use of the steady-state approximation.
In our experiments, we measure intestinal abundances in di-associations, P II ,
and mono-associations, P I . The carrying capacity, Ki, is unknown. We assume
Ki can be well approximated by the average P
I
i across fish, i.e. that the average
abundance of a species in the absence of other species is similar to its carrying
capacity in the gut. Therefore:
II
≈ Pi − ⟨P
I
i ⟩Cij (2.8)
P IIj
where the angle brackets indicate the mean across fish. Eq. 3.3 provides a simple
characterization of inter-species interactions in the context of the standard
competitive Lotka-Volterra equations.
As discussed in the main text, there is no reason the intestinal microbial
populations must be governed by Lotka-Volterra dynamics or any linear interaction
model. We therefore consider a model of logarithmic interactions and a more
general family of power-law models. The logarithmic model, influenced by the
order-of-magnitude changes in species abundance induced by competition, can
follow from a phenomenological recasting of the competitive Lotka-Volterra model
in terms of log-transformed populations:
45
( )
d log10 Pi − log10 Pi − Cij log= f(P ) 1 10 Pji , (2.9)
dt log10Ki
where we note that the first factor on the right hand side can be any generic
function of Pi. Again assuming steady state dynamics and estimating log10Ki by
the average of log10 Pi, the interaction coefficients are given by:
≈ log P
II I
C 10 i
− ⟨log10 Pi ⟩
ij . (2.10)
log P II10 j
which is Eq. 2.6 of the main text. We note that regardless of the validity
of the dynamical model, Eq. 2.9 can be considered as a phenomenological
characterization in which interaction effects are logarithmically additive (i.e.
multiplicative) in population size.
Finally, we consider interactions with a power-law dependence on the size of
the influencing species:
( )
dPi − Pi − Cij(P )
α
j
= f(Pi) 1 , (2.11)
dt Ki
from which
II I
Cij ≈
Pi − ⟨Pi ⟩ (2.12)
(P II)αj
As noted in the main text, we consider sub-linear (α < 1) and super-linear
(α > 1) interactions. Linear interactions (α = 1) are equivalent to the standard
Lotka-Volterra model.
46
2.4.7.2. Pairwise Interactions in the Presence of Additional Species
To compare interactions between communities of two species and
communities of more species, namely four or five, we extended the above
parameterization. The standard Lotka-Volterra dynamical model for an arbitrary
number of species is given by
 ∑  Pi − CijPjdPi − j ̸=i = riPi 1  (2.13)
dt Ki
where the sum runs over species. Again considering steady-state,
∑
Pi = Ki + CijPj (2.14)
j≠ i
As discussed in the main text, we infer CVij by comparing the abundance of
species i in fish co-inoculated by a set of species excluding and including species j.
Again denoting the number of species by a superscript, subtracting P IVi from P
V
i
gives
P V − P IV = CV P Vi i ij j , (2.15)
from which
V
V Pi − P IVC iij = . (2.16)P Vj
Similar expressions follow for the multi-species interaction coefficient in the
log-transformed model:
V
V log10 Pi − log P IVC 10 iij = . (2.17)log V10 Pj
47
and the power-law model:
V
V Pi − P IVC = iij . (2.18)(P V )αj
The interaction matrices for calculated for the two-species experiments (CIIij ),
using Eq. 2.10 and four-to-five-species experiments (CVij ), using Eq. 2.17 are shown
in Fig S2. As noted in the main text, we found similar patterns distinguishing
CIIij and C
V
ij , namely weaker and more positive interactions in the latter case,
throughout the full range of α examined (Fig 2.3, A.4).
2.4.8. Sampling and Parameter Estimation
For each di-association, we measured the species abundances P IIi and P
II
j in
N=10 to 30 fish and calculated CIIij for each according to Eq. 2.8, 2.10, 2.12 above.
To ensure that the measured abundance changes can be attributed to the presence
of the second species we ignored data for which either P IIi or P
II
j was zero. Each
set of abundance measurements provides a Cij value via the above equations.
To determine the most likely Cij from this set, as well as the uncertainty in our
estimation of Cij, we make use of sampling. With 3000 repetitions, we randomly
sampled 75% of each dataset without replacement and calculated the mean and
standard deviation of the Cij.
To determine the number of repetitions needed to accurately estimate Cij,
we performed the same analysis on simulated population data, drawing P IIi from a
log-normal distribution, using a fixed ⟨log I10 Pi ⟩, and drawing Cij from a normal
distribution of known mean and standard deviation. Using these we compute
log II10 Pj using Eq. 2.10, and from the simulated data we calculated the mean and
standard deviation of the interaction coefficient as above. We found that sampling
75% of the dataset, ∼ 2000 repetitions yield the standard deviation of Cij to an
48
accuracy of 90%. The resulting mean and standard deviation of CIIij , one for each
species pair, are shown in the figures in the main text and Fig A.2A.
Interaction parameters in multi-species experiments, i.e. CVij , were similarly
calculated from subsampling abundance data from both the four species and the
five species experiments. The mean and standard deviation of the CVij coefficients
calculated using this model are shown in Fig S2B. As in the calculation of the di-
association-derived CIIij , we considered only data from fish in which each of the
species inoculated in the four and five species experiments was detected, ensuring
that the effects from all other species are consistent between experiments.
In vitro interaction coefficients were also calculated using Eq. 2.10 from
abundance measurements in pairwise competition experiments in Lysogeny broth
(LB) medium. The parameters shown in Fig A.3 reflect the mean and standard
deviation of the coefficients from six replicates for each of the pairs of species.
2.4.9. Relative Abundance Correlations
As noted in the main text, the Pearson correlation coefficients between
the relative abundances of species are commonly interpreted as measures of
inter-species interactions. Denoting the relative abundance of species i (i.e. the
population normalized to the sum of all species populations) in fish n as (pi)n, we
calculate the correlation coefficient as usual:
∑N ( )( )
rij = √ (p ) − ⟨p ⟩ (p ) − ⟨p ⟩∑n=N (
i n
1 )i j n j (2.19)2 ∑N ( )2
(pi)n − ⟨pi⟩ (pj)n − ⟨pj⟩
n=1 n=1
49
where N is the total number of fish examined. Note that by construction, rij is
symmetric.
2.4.10. Predicted Five-Species Abundance Distributions
We asked whether we can recreate the experimentally observed five-species
abundance distributions using only the pairwise interaction coefficients from
Section 2.4.7.1. Focusing first on the logarithmically additive model, we construct
a predicted abundance value for each species by linearly combining the pairwise
effects of each of the other species:
∑
log P V10 i = ⟨log10 P Ii ⟩+ CII Vij log10 Pj . (2.20)
j ̸=i
P Vi and P
V
j are the predicted abundances of species i and j respectively.
This gives, for the five bacterial species in our system, five equations with five
unknowns, namely the predicted Pi. To determine the distributions of predicted
abundances for each of the five species, we randomly sampled each CIIij from the
interaction coefficient distributions obtained in Section 2.4.7.1 . We performed this
sampling 250 times to generate a comparable number of points to the experimental
distribution (N=202). Both the predicted and the measured distributions are
shown in Fig 2.4D. To mimic the experimental detection limit of approximately
25 cells, all predicted abundances below 25 were set to zero.
Similarly, we calculated predicted abundance distributions for using the
linearly additive model in absolute abundance (i.e. competitive Lotka-Volterra),
for which ∑
P V = ⟨P Ii i ⟩+ C VijPj . (2.21)
j ̸=i
50
As above, we generated predicted distributions for each of the five species, shown
in Fig A.6.
As discussed in the main text, the predicted distributions from both
pairwise-additive models indicate that the five species are less likely to coexist
than is observed experimentally.
2.4.11. Five-Species Shannon Entropy
We calculated the Shannon entropy, a metric commonly used to measure
species richness in multi-species communities. This is defined as:
∑
H = − pi ln pi (2.22)
i
where pi is the relative abundance of species i. For a neutral (non-interacting)
community of N species, pi = 1/N , on average. For five species therefore, a neutral
model gives H = ln(5) = 1.61. To calculate H from measured data, we used
relative abundances of each bacteria species and calculated the Shannon entropy
using Eq. 2.22. The mean Shannon entropy, averaged over all fish, is reported in
the Discussion section of this chapter.
51
CHAPTER III
SPATIAL STRUCTURE AND INTERACTION MECHANISMS
This work is adapted from co-authored work currently under revision by D.
Sundarraman, T. J. Smith, , J. V. Z. Kast, K. Guillemin and R. Parthasarathy,
Disaggregation as an interaction mechanism among intestinal bacteria, submitted
to Biophysical Journal (2022). I was involved in the experiments, analysis and
writing of this work.
3.1. Introduction
The vertebrate gut is home to a diverse set of interacting microbial species,
with microbiome composition correlating with aspects of host metabolism,
digestion, immune response, development, and more [12–26], Dysbiosis of the gut
microbiota has been linked to several diseases such as obesity, IBD and rheumatoid
arthritis [2–11]. Understanding the forces that shape intestinal communities is
therefore of considerable importance, and many studies have explored how factors
such as host diet and genetics can affect microbiome assembly and function. In
general, spatial organization is a common and important aspect of microbial
ecosystems; surface attached biofilms, microbial mats, and marine snow provide
well-known examples [79, 149–152]. This likely holds in the gut as well, but
spatial information is usually difficult to obtain. Conventional approaches based
on, for example, metagenomic sequencing of fecal samples are blind to structure,
and most spatially resolved methods typically involve coarse sampling or fixation
methods that can drastically perturb the features present in live animals [82].
Regarding inter-microbial interactions in particular, we know little about the
52
mechanisms at play, which may include segregation to enable coexistence, contact-
mediated competition, or more complex spatially-varying signals.
Larval zebrafish provide a model vertebrate system in which to observe
spatial relationships among gut microbial species. The animals’ optical
transparency and amenability to gnotobiotic techniques [47–49, 122], together
with a library of fluorescently tagged native gut bacterial isolates [50], allows
controlled experiments involving live imaging of bacterial communities [51–
53, 92, 106]. Mapping the spatial structure of a number of gut bacterial species
established that each, in mono-association with its host, has a characteristic
fraction of its population forming dense, three-dimensional aggregates and is
preferentially localized in particular regions of the gut [91]. Given the well-
established ecological principle of competitive exclusion, by which species
competing for the same limited resource cannot stably coexist, we might expect
that if introduced together, two species with an affinity for the same region will
strongly compete, leading to greatly reduced abundance for at least one of the
two relative to the mono-association value. Spatial segregation of competitors,
whether static or dynamic, provides a route to co-existence [153–155]; thus we
might expect that reducing spatial overlap of intestinal species should weaken
competition. How these principles may or may not be manifested in a real
vertebrate gut remains unknown, and as we show below, an example with a pair
of zebrafish-native bacterial species confounds simple expectations.
Inter-species interactions are also influenced by timing, and the outcomes of
competition can vary depending on whether species arrive simultaneously or not
to an environment. Various ecological studies have shown that early colonizers
can have an advantage over later species and can often resist invasion by late
53
competitors [55, 156, 157]. Regarding the gut microbiome, the pre-existing
community is believed to contribute to resistance to pathogen invasion [38, 158],
though a greater number of controlled studies in animal models are needed to
uncover underlying mechanisms [37, 158].
To illuminate possible connections between spatial organization and inter-
bacterial interactions in the context of a vertebrate intestine, we performed a set
of live-imaging-based studies of a strongly interacting bacterial species pair, using
light sheet fluorescence microscopy to visualize bacterial locations and dynamics
in situ. Prior work based on measuring species abundances, without spatial
information, characterized the interactions between various zebrafish-commensal
gut bacterial species and found several examples of strong competitive pairwise
interactions, one such pair being Aeromonas ZOR0001 and Enterobacter ZOR0014
[159], each of which form dense aggregates in mono-association [51, 53, 91]. These
and other strong pairwise interactions were dampened in the presence of additional
bacterial species [159].
As described here, imaging reveals that both of these species colocalize
and co-aggregate in the larval midgut. We introduce a derivative of Aeromonas
ZOR0001, Aeromonas-MB4 that is deficient in biofilm formation in vitro in the
presence of a glycan commonly found in the gut and find that in vivo, it is largely
planktonic and motile. The dispersed strain leads to an even lower Enterobacter
abundance than the co-localized wild type, and induces a striking dissociation of
Enterobacter aggregates. In the presence of an already established community of
multiple bacterial species, competition and disaggregation effects are diminished,
coexistence is enhanced, and the planktonic isolate behaves more similarly to the
wild-type parental strain. Overall, our observations point to a surprising and
54
unanticipated mechanism by which spatial structure can mediate gut bacterial
interactions, namely induced disintegration of aggregates.
3.2. Results
3.2.1. Aeromonas-MB4 exhibits planktonic behavior in vitro and in
vivo.
We consider two strains of Aeromonas. One is a zebrafish gut commensal,
Aeromonas ZOR0001 (referred to as AE). The other is an Aeromonas isolate
(referred to as AE-MB4), derived through directed evolution of AE in vitro. The
zebrafish-native AE forms aggregates in vitro in certain growth media, including
sterile embryo medium supplanted with 0.4% N-Acetylglucosamine (GlcNAc), a
sugar prevalent in intestinal mucus [93, 160–162]. The AE-MB4 strain was derived
from AE by repeated passaging of fractions of such media that were devoid of
visible aggregates, thereby selecting for bacteria deficient in aggregation in the
presence of GlcNAc (Fig A.7). The molecular genetic characterization of AE-MB4
is described elsewhere [163]; we focus here on the spatial dynamics of this strain
and the consequences of this altered aggregation behavior on community structure.
55
FIGURE 3.1. (A) Bright field image of a 7dpf zebrafish with the gut outlined
in red. Bar: 500µm (B & C) Maximum intensity projections of 3D images of
the intestine of zebrafish mono-associated showing (B) zebrafish-commensal
AE bacteria (orange), which forms large midgut aggregates and (C) AE-MB4
(magenta), which is largely planktonic. Dotted curves roughly indicate the gut
boundary. Bar: 100µm. (D) Mean normalized intensity profiles of AE (orange) and
AE-MB4 (magenta) along the anterior-posterior axis. In contrast to AE, which
localizes predominantly in the midgut, much of the AE-MB4 population is in the
anterior bulb. Circles indicate the center of mass. Curves are averaged from N =
6 fish each. (E) The probability of a cell being in an n-cell cluster for AE (orange)
and AE-MB4 (magenta), mono-associated with zebrafish. Circles and error bars
indicate the mean and uncertainties, calculated using jack-knife resampling. Tick
marks indicate bin intervals, e.g. the orange and magenta bars between 0 and
100 correspond to AE and AE-MB4 probabilities to be in clusters of size n = (0-
1]. Data are from N = 6 and 8 fish for AE and AE-MB4, respectively. (F) The
cumulative cluster size distributions (p[cluster size ≥ n]) for AE (orange) and AE-
MB4 (magenta) in mono-association calculated from all clusters found in N = 6
and 8 fish, respectively, showing power law behavior for n < 102. The dashed lines
illustrate the slopes from power law fits in the range n = 1 to 102, with exponents
0.8 for each strain. At large n, the AE distribution plateaus, reflecting the presence
of large clusters.
56
FIGURE 3.1.
57
We first characterized the spatial distribution of bacteria in initially germ-
free zebrafish larvae (Fig 3.1A) mono-associated with either the commensal AE or
the AE-MB4 isolate. We mono-associated initially germ-free zebrafish at 5 days
post-fertilization (dpf) with each strain and imaged 48 hours later using light
sheet fluorescence microscopy (Methods). Wild type AE forms dense clusters in
the midgut (Fig 3.1B, D), as seen in prior work [51–53, 91, 92], whereas AE-MB4
resides primarily in the anterior with a large fraction of the population in the form
of discrete planktonic cells (Fig 3.1C,D). Image analysis (see Methods) identifies
individual bacterial cells and aggregated clusters and provides estimates of cluster
populations. Cluster size statistics highlight the differences between the strains.
After pooling measured cluster sizes from all fish, we compute the probability
of a cell being in a cluster of a given size i.e. the ratio of the total number of
cells found in a given cluster size bin to the net population from all clusters. For
example, the bar associated with bin (0− 100] reflects the frequency of finding cells
in clusters of size 1 in the total population. We plot this measured frequency for
different cluster size bins in Fig 3.1E. AE cells are most likely to reside in large
clusters (n = 103.5), a population completely absent in AE-MB4, and AE-MB4
cells are over 30 times more likely to be planktonic than AE (Fig 3.1E). All cluster
sizes are provided as supplemental Material. We also plot the reverse cumulative
probability distribution of cluster sizes, p(n), i.e. the probability that a given
cluster is composed of more than n bacterial cells (Fig 3.1F). The distributions for
both species have characteristic power law form for small clusters (n < 100) with
a slope of m = 0.8 ± 0.1 [92]. The disparity in p(n) between AE and AE-MB4
grows as n increases and is especially pronounced for clusters of size n > 1000 as
the distribution of AE becomes shallower. Earlier work studying intestinal cluster
58
size distributions in a range of species has shown that characteristic power-law
features arise from fragmentation and growth dynamics, and a shallow decline of
p(n) at large n is indicative of aggregation [92], consistent with observations here.
3.2.2. A highly aggregated commensal species undergoes rapid
fragmentation in the presence of the Aeromonas-MB4 isolate.
To study the role of spatial coincidence in competition in the zebrafish
gut, we studied interactions of Aeromonas and Aeromonas-MB4 with a different
zebrafish-commensal bacterial species, Enterobacter ZOR0014 (referred to as
EN). Past work based on abundance measurements established that the parental
AE strain has strong negative interactions with EN, with the population of the
latter reduced by over an order of magnitude from its mono-association value
when di-associated with AE (Fig 3.2E) [159]. EN in mono-association is almost
completely aggregated, with only about 10% of its population in the form of
planktonic individuals, forming large clusters primarily in the midgut region of
the intestine (Fig 3.2A) [53, 91]. Consistent with earlier measurements, its cluster
size cumulative probability distribution and the probability of a cell being in a
cluster of a given size reflect the highly aggregated spatial character of the species
(Fig A.8,A.9).
59
FIGURE 3.2. (A-D) Maximum intensity projections of 3D images showing (A) EN
(green) in mono-association forming large clusters in the midgut, (B) EN (green)
in di-association with AE (orange) showing co-aggregation of both species in the
midgut, (C) EN (green) in di-association with AE-MB4 (magenta) showing sparse
EN amid abundant planktonic AE-MB4 cells, and (D) EN (green) after invasion
by AE-MB4 (magenta) showing large populations of single cells and small clusters
of EN throughout the gut. Inset: Many single cells of EN are evident. Dotted
curves roughly indicate the gut boundaries. Bar (A-D): 100µm. (E) Twin axis
plot showing the abundance (left axis) in log10(CFUs/gut + 1) and extinction
fraction (right axis) for EN (green), AE (orange) and AE-MB4 (magenta) in mono-
association (circles), di-association (Xs), and when EN is invaded by AE-MB4
(diamonds). Each marker indicates abundance data from a single fish. The blue
line markers and black circular markers depict mean and standard deviation of the
extinction fraction and abundance for EN and AE-MB4 in each of the experiments
respectively. From left to right, N = 15, 12, 17, 27, 27, 36, 36, 23, and 23 fish. (F)
The planktonic fraction per fish for EN in (from left to right) mono-association,
dissociation with AE, dissociation with AE-MB4 and when invaded by AE-MB4.
From left to right, N = 7, 10, 12, 12 fish. G. As in Fig. 3.1E, The probability of a
cell being in an n-cell cluster for EN (i) in mono-association, (ii) in di-association
with AE, (iii) in di-association with AE-MB4, and (iv) invaded by AE-MB4.
Circles and error bars indicate the mean and uncertainties. Data are from N =
7, 10, 12, and 12 for (i)-(iv), respectively.
60
FIGURE 3.2.
61
We performed di-association experiments (as in [159]) in which both AE
and EN were inoculated simultaneously at 5 dpf and the spatial distribution was
assessed at 7 dpf. Through imaging both species in di-association, we found that
AE and EN often co-aggregate in the same region of the intestine (Fig 3.2B),
suggesting that strong competition could be a consequence of both species having
an affinity for similar locations. These co-aggregates of AE and EN consist of
intermingled clonal clusters of each species, not well-mixed individuals of each
species.
We next investigated the consequences of the AE-MB4 traits on the
Aeromonas-Enterobacter competition. In di-association with AE-MB4, EN exhibits
almost an order of magnitude lower mean abundance (2.2 ± 0.7 ) than with AE
(3.0 ± 0.6), and roughly two orders of magnitude lower than its mono-association
value (4.4 ± 0.7) (Fig 3.2E), in contrast to expectations that the species’ dissimilar
localizations may attenuate competition. We also note that extinction of EN is
more frequent in di-association with AE-MB4 (extinction fraction = 0.2 ± 0.2
), than with AE (extinction fraction = 0.1 ± 0.1 ). The effect on the spatial
distribution of EN is striking; nearly all EN are found as discrete individuals
(Fig 3.2C) , with the planktonic fraction being 0.9 ± 0.2 (mean ± std. dev.)
compared to 0.1 ± 0.1 in mono-association. EN cells are never found in clusters
of size n > 102, the most common range in mono-association, and in nearly 80% of
the fish examined, EN shows no apparent clustering at all. The individual cells of
EN rarely co-aggregate with AE-MB4 clusters, in contrast to frequently occurring
co-aggregates of both species in the di-association of EN with AE.
To test how priority effects could impact interactions, we examined the
abundance and spatial distribution of EN after invasion by either the AE or AE-
62
MB4 strains. Prior work established that EN reaches steady state abundance
in approximately 8-12 hours [51, 53]. Thus, to ensure a stable pre-invasion EN
population, we inoculated the second species 24 hours after the introduction of
EN. When challenged with AE, the initial advantage serves to benefit EN, with
most of the fish sustaining EN aggregates (Fig A.10). We anticipated priority
effects to similarly impact invasion by AE-MB4. Contrary to this expectation,
we found that EN is unable to persist as large aggregates in the gut (Fig 3.2D).
Instead, the population consists of a large fraction of individual cells and some
small clusters with a mean planktonic fraction of 0.7 ± 0.3 (Fig 3.2F,G), similar to
that observed in the di-association experiments. The remaining clusters of EN in
these experiments are often fragmented and sparse, unlike the dense aggregates of
EN seen in other experiments.
Through time series imaging of EN-colonized fish starting 6 hrs after invasion
with AE-MB4, we observed striking, rapid fragmentation of large EN aggregates
into individual cells and small clusters within timescales of hours (Fig 3.3A,B).
Such disintegration has not been observed in mono-association or di-association
with AE and thus we attribute it to the presence of AE-MB4. The planktonic EN
showed little motility. Analyzing the rate of change of isolated sub-populations
identified in images as planktonic and aggregated after fragmentation of the EN
aggregate, we found the growth rate (r) of individual cells (r = 0.24 ± 0.11 hr−1,
mean ± std. dev. for N=3 fish) to be approximately three times lower than that of
clusters (r = 0.66 ± 0.23 hr−1 for N=4 fish) (Fig 3.3C). This was calculated from
exponential fits of the abundance over time after identification of bacterial clusters
and individuals in consecutive 3D images (Methods). This measurement takes
into consideration regions where there is only growth, not influx or expulsion. The
63
large planktonic fraction together with the lower r of the planktonic relative to the
aggregated form implies a reduced ability of EN to persist in the gut following AE-
MB4 invasion. The growth rate of AE-MB4 in the invasion experiments (r = 0.76
± 0.04 hr−1 for N=3 fish) was found to be higher than that of EN, consistent with
the expectation that a fast growing species outcompetes a slow growing species.
To better understand the implications of a high planktonic fraction of usually
aggregated EN in the gut, we simulated the dynamics of clusters in the zebrafish
intestine undergoing four key processes, growth, aggregation, fragmentation
and expulsion. Simulation of these processes has been shown to reproduce
characteristic features of cluster size distributions observed in experimental data
[92].
A detailed description of the stochastic model and its parameters is
provided in Methods. In brief, growth of each cluster is logistic and the total
abundance (sum of all clusters) is bound by a carrying capacity, K, equal to the
mono-association abundance of the species and growth rates of individuals and
clusters (rpl and ragg) being the experimentally measured parameters (Fig 3.3C).
Aggregation or the combining of two clusters of size n, m to form a single
cluster(n+m) occurs at an aggregation rate α = α (n*m)1/3nm where the exponent
1/3 signifies collision likelihood proportional to cluster diameter. The value of α is
fixed at log10(α) = −2.5. Similar to aggregation, a cluster of size n undergoes size-
dependent fragmentation of single cells at a rate β = (n)2/3, corresponding to the
breakup of a single cell from a cluster of size n with the 2/3 exponent reflecting
that only cells on the surface have a likelihood of breaking up from the cluster. β
is varied from 0.02 hr−1 to 1 hr−1 in steps of 0.025 hr−1. For a cluster of size 1000,
this range implies 2-100 cells fragmenting per hour. Size-independent expulsion of
64
aggregated clusters is fixed at a rate corroborated by previous studies, λagg=0.1
hr−1 while the expulsion rate of planktonic cells is varied logarithmically between λ
= 0.025 and 10 hr−1pl .
We combine these processes in a stochastic simulation, with one of
fragmentation, aggregation or expulsion reactions occurring at a given time
step determined by the rates for each reaction computed for each cluster at the
given time step. First, the simulation is run to determine the steady state mono-
association cluster distribution for fixed α, β and λ after 24 hrs. After generating
this initial EN distribution, we simulate the AE-MB4 invasion of EN aggregates
by varying β. We run this for the full range of parameter values of λpl stated
earlier. Averaging total abundance from 500 replicates for each pair of parameters
λ
of expulsion rate ratio pl and fragmentation rate β we obtain the abundance
λagg
phase diagram shown in (Fig 3.3D). When λpl < λagg , we find that abundance
corresponds to the mono-association value, irrespective of changing fragmentation
rates. For λpl > λagg, the abundance corresponds to our experimentally observed
abundance values for an expulsion rate 0.13 hr−1 < λpl < 2.5 hr
−1 with the overlap
more distinct in the high fragmentation regime. This is physically plausible as non-
motile cells undergo a more constant efflux in the intestine, being more susceptible
to flows compared to large aggregates that are removed in mass-expulsion events,
also observed experimentally. This suggests that the properties of planktonic non-
motile EN, with its reduced growth rate and its increased susceptibility to being
expelled from the gut, can explain the low experimentally observed values of EN
abundance.
65
FIGURE 3.3. (A) Maximum intensity projections of the midgut region from a
time series showing an aggregate of EN that fragments following invasion by AE-
MB4. Times indicate hours post-invasion. Dotted curves roughly indicate the gut
boundary. Bar : 50 µm. (B) The planktonic fraction of the EN population over
time for 3 fish measured starting 6 hrs. after invasion by AE-MB4.(C) The number
of cells normalized by the initial value over time for aggregated (blue, dotted) and
planktonic (teal, solid) EN populations after invasion by AE-MB4. For aggregated
populations, each dataset depicts the growth of a single cluster of EN in a distinct
fish (N=4). For planktonic populations, each dataset corresponds to a region
comprising only planktonic cells in a distinct fish (N=3). The black curves indicate
the mean growth rates: m = 0.66 ± 0.23 hr−1 and 0.24 ± 0.11 hr−1 for aggregated
and planktonic cells, respectively. (D) Simulated abundance of EN as a function of
λ
the ratio of the expulsion rates for planktonic to aggregated populations ( pl ) and
λagg
the fragmentation rate β (hr−1). Experimentally measured parameters were used
for the growth rates of aggregated (ragg) and planktonic (rpl) EN clusters (shown
in C). The shaded region corresponds to the experimentally measured abundance
(log10(abundance) = 2.0 ± 0.9). See the Methods section for details.(E) Brightfield
image of EN-AE-MB4 co-aggregates in 0.4 % GlcNAc solution. Scale bar: 1mm.
(ii) Single plane of a 3D image showing co-aggregated populations of AE-MB4
(magenta), EN (green) and overlapping populations (white) in 0.4 % GlcNAc.
Scale bar: 10µm.
66
FIGURE 3.3.
67
We study the aggregation of EN in the presence of AE-MB4 also in in vitro
experiments, comprising 0.4 % GlcNAc solution, a medium in which AE-MB4
resists aggregation (Fig A.7). Strikingly, in co-culture experiments of AE-MB4 and
EN in the same medium, AE-MB4 readily co-aggregates with EN and we observe
large clusters comprising both species Fig. 3.3E. These observations show that in
vivo interaction dynamics and outcome are markedly different from those in vitro.
3.2.3. The Aeromonas-MB4 isolate displaces the closely related
parental AE strain
To test whether the strong competition effects and dynamics observed are
specific to EN and to examine how other gut bacterial species may be impacted
by AE-MB4, we studied AE-MB4’s interactions with the parental zebrafish isolate
AE strain. We examined intra-species interactions by inoculating, as in previous
experiments, initially germ-free zebrafish with fluorescently labeled AE at 5 dpf
and challenging by introduction of either AE-MB4 or differently-labeled AE 24
hours later. The abundance of AE after invasion by AE-MB4 (log10(abundance) =
2.8 ± 0.5) drops by over an order of magnitude compared to its mono-association
value (log10(abundance) = 4.2 ± 0.4). Comparing the probabilities of finding AE
cells in different sized clusters when invaded by itself versus when invaded with the
planktonic AE-MB4, we discovered an absence of very large clusters (n = 104) in
the presence of AE-MB4 (Fig 3.4), suggesting that the AE-MB4 isolate impacts
the parental bacteria’s ability to either form or maintain dense aggregates. The
dynamics of interactions differ, in that we observe no dissociation of AE aggregates
and instead expulsion of large AE clusters with often only small clumps persisting.
68
FIGURE 3.4. (A & B) Maximum intensity projections from zebrafish initially
colonized by AE (orange), invaded 24 hrs. later by (A) differently labeled but
otherwise isogenic AE (blue) (B) AE-MB4 (magenta). Images (A) and (B) were
taken approximately 24 hrs. post-invasion. Dotted curves roughly indicate the
gut boundary. Bar: 100µm. Inset to (A): a bacterial aggregate comprising both
populations of AE (white overlapping regions). (C) Cluster probabilities as in
Fig. 3.1E for AE (orange) when (i) invaded by others of the same strain (AE)
and (ii) invaded by AE-MB4, both after 24 hrs. Circles and error bars indicate
the mean and uncertainties. Data are from N = 8 and 10 fish for (i) and (ii),
respectively.
3.2.4. Multi-species communities dampen strong AE-MB4-EN
interactions
We next studied the interactions of EN and AE-MB4 in the presence of
additional species, using a set of zebrafish commensal bacteria whose interactions
have previously been characterized [159]. The additional gut community
members comprised Plesiomonas ZOR0011 (referred to as PL), Acinetobacter
ZOR0008 (AC), Pseudomonas (ZWU0006) (PS), and Aeromonas (AE). First,
we co-inoculated four species (EN, AC, PL and PS) at 5 dpf and observed high
69
abundances and large populations of aggregated EN (Fig 3.5A) in the midgut
region of the intestine, resembling its mono-association form (Fig 3.2A, E, G),
suggesting that these species have little effect on Enterobacter’s spatial structure.
Then we examined the impact of AE-MB4 when also co-inoculated along with
the other four species and found that on average populations of EN were more
planktonic (planktonic fraction = 0.7 ± 0.2), (Fig 3.5B), as in the di-association
experiments. We also noted frequent extinction of EN, with approximately 1/3
of the fish showing no trace of EN. Overall, however, EN abundance was slightly
higher (log10(abundance) = 2.8 ± 0.5) than in the di-association experiments
(log10(abundance) = 2.3 ± 0.7) (Fig 3.5E,G). This suggested that the negative
AE-MB4 interaction effects are still present but relatively weaker in these five
species co-inoculation experiments. The abundance data (Fig 3.5E) includes fish
in which all species inoculated in the experiment were detected and all species
excluding PL were detected. PL generally showed rare colonization across all
datasets.
70
FIGURE 3.5. Maximum intensity projections of a larval zebrafish gut showing
the spatial distribution of (A) EN (green) in the presence of four other commensal
gut bacterial species forming large aggregates in the intestine (B) EN (green)
in the five-species co-inoculation experiment (including AE-MB4 (magenta))
has smaller fragmented clusters and many single cells (C) EN (green) in the
experiment when five species (including AE) are invaded by AE-MB4 (magenta)
where EN forms larger clusters with some single cells and AE-MB4 populations
show some aggregation. The dotted curve roughly defines the gut wall. Bar:100µm
(D) Mean normalized intensity profiles measured along the anterior-posterior
axis showing the spatial distribution of EN (green) and AE-MB4 (magenta)
along the intestine when (i) EN is invaded by AE-MB4 and (ii) all five species
are invaded by AE-MB4 showing large aggregates of EN persist in the midgut.
The circular markers indicate the center of mass. Data are from N = 13 and 11
fish for (i) and (ii) respectively. (E) Twin axis plot showing the abundance of
EN (left axis) in log10(CFUs/gut + 1) and extinction fraction (right axis) when
in mono-association, in di-association with AE-MB4, invaded by AE-MB4, in
the four-species co-inoculation experiment, in the five species co-inoculation
experiment (including AE-MB4), in the five species co-inoculation experiment
(including AE) invaded by AE-MB4. Abundances from individual fish are depicted
with green circular markers. Mean and standard deviation for abundance and
extinction fraction of EN are shown with black circular and blue line markers
respectively. N (from left to right) = 15, 36, 23, 21,11, 87, 19 fish. (F) The
planktonic fraction per fish for EN when invaded by AE-MB4, in the four–species
co-inoculation experiment, in the five-species co-inoculation experiment (including
AE-MB4) and when the five-species (including AE) are invaded by AE-MB4.
Black markers indicate the mean and standard deviation. N (from left to right)
= [12, 5, 2, 14] fish. (G) The probability of being in an n-cell cluster of EN when
(i) invaded by AE-MB4 (ii) in the four-species co-inoculation experiment (iii) in
the five-species co-inoculation experiment (including AE-MB4) (iv) in the five-
species co-inoculation experiment (including AE) invaded by AE-MB4. Mean and
uncertainties determined from jack-knife resampling are shown with black markers.
N (top to bottom) = [12, 5, 2, 14] fish.
71
FIGURE 3.5.
72
To determine whether these interactions persist in spite of priority effects,
we performed invasion experiments with fish co-inoculated with all five species
(AC, AE, EN, PL and PS), similar to mono-association EN experiments. We
first studied the abundance distribution of EN in the control experiment, when
no invader was introduced. We found that EN is more likely to be aggregated in
these fish, resembling its typical structure with mean abundance in these fish =
103.0±0.8 (Fig 3.5E).
On challenging this multi-species community with AE-MB4, we found
strikingly different results from that of the mono-association invasion. Large
aggregates of EN were able to persist in the presence of other species and the
mean planktonic fraction of EN (0.2 ± 0.3) was lower than in the mono-association
invasion experiments (0.7 ± 0.3) (Fig 3.5 D,C and F). The likelihood of a cell
being in clusters of size n > 103 goes up from 67% in the mono-association invasion
experiment to 97% in the multi-species challenge experiments (Fig 3.5G), while
the planktonic probability is reduced from 25% to 2%. In time series experiments
observing the dynamics of EN aggregates in the presence of other species, large
aggregates of EN either resist breakup or undergo slower fragmentation relative
to the fast timescales observed in single species mono-association invasion
experiments. Together, all of these observations suggest that having a diverse and
stable pre-existing community can inhibit strong interaction effects due to the
invader.
To quantify the strength of interactions in different contexts, we calculated
an interaction coefficient metric measuring the interaction effect of species j on
species i and defined as follows as in [159]:
73
1−2 ⟨log I IIC = 10 Pi ⟩ − log10 Piij (3.1)log10 P IIj
Here, P Ii is the mono-association abundance of species i, P
II and P IIi j are the
abundances of species i and j in di-association. This follows from a competitive
Lotka-Volterra model using log-transformed abundances in place of absolute
abundance, described in detail in [159].
One can extend this interaction coefficient to quantify the effect of a single
species j on another species i in the five species experiments:
log IV V
C4−5 10
Pi − log10 Pi
ij = (3.2)log10 P
V
j
These equations can be applied to the invasion experiments where P IIi and
P IIj are replaced by the abundances of species i and j in the invasion experiments.
Similarly, Eq.3.2 translated to the five species invasion case can be written as:
5−6 log P
V
10 i − log10 P V IC = iij (3.3)log10 P V Ij
Where P Vi is the abundance of species i in the five species (AC, AE, EN,
PL and PS) co-inoculation experiment, P V Ii is the population of i post-invasion
with species j and P V Ij is the abundance of the invader j. Using equations Eq. 3.1,
3.2 and 3.3, with the experimental abundance data, we can extract the mean and
uncertainty of the interaction coefficients for various experiments. We measured
the interaction coefficients in the di-association and invasion experiments to be
similar with C1−2ij = −0.64 ± 0.03 for the di-association and −0.65 ± 0.03 for the
invasion experiments. These signify strong negative interactions with populations
of EN dropping by two orders of magnitude relative to mono-association. In the 5
74
species co-inoculation experiments with AE-MB4, we found C4−5ij = −0.11 ± 0.10,
significantly weaker than C1−2ij for di-association. For the invasion experiments
we found the magnitude of interactions to be zero within uncertainties C5−6ij =
−0.02 ± 0.06. These coefficients affirm implications of previous work that higher
order interactions dampen strong pairwise effects and promote co-existence in a
diverse community.
To further examine the nature of interactions in a multi-species context
and how other species are impacted in the presence of AE-MB4, we returned
to characterizing the effects of AE-MB4 on the parental AE strain. In mono-
association, challenged by AE-MB4, AE was unable to persist in clusters of n
≥ 104 cells with no likelihood of being in such clusters (Fig 3.4C (ii)). When
additional species are present, we see large aggregated populations of AE in the
midgut and cells are 53% likely to be found in aggregates of size ≥ 104 cells
(Fig 3.6A, B, C). As with EN, greater gut community diversity counteracts
the changes induced by the AE-MB4 isolate. Surprisingly, we also found
that AE-MB4, highly planktonic on its own, forms aggregates in the invasion
experiment with other species present, also altering its localization in the intestine
((Fig 3.6B(ii)D).
75
FIGURE 3.6. (A) Maximum intensity projection of a larval zebrafish gut showing
a large cluster of AE (orange) after invasion by AE-MB4 (magenta) in the presence
of the other four species. Small clumps of AE-MB4 are interspersed in the AE
clusters. Bar: 100µm (B) Mean normalized intensity profiles measured along the
anterior-posterior axis showing the spatial distribution of AE (orange) and AE-
MB4 (magenta) along the intestine when (i) AE and (ii) all five species (including
AE) are invaded by AE-MB4. The circular marker indicates the center of mass.
Data for (i) and (ii) are from N = 10 and 11 fish respectively.(C) The probability
of being in an n-cell cluster for AE when (i) invaded by AE-MB4 (ii) invaded
by AE-MB4 in the presence of the other four species. The mean and standard
deviation are indicated with circular markers. The x tick marks indicate the
bin intervals i.e. the bar between the tick marks 0 and 100 corresponds to the
probability of being in a cluster within (0-1]. Data are from N = 10 and 14 fish for
(i) and (ii) respectively. (D) The probabilities (as in C) for AE-MB4 (i) in mono-
association (ii) after invading the five species. The mean and standard deviation
are indicated with circular markers. N = 8 and 11 fish for (i) and (ii) respectively.
3.3. Discussion
We report an unexpected mechanism of interactions among gut bacterial
species: induced disaggregation. Specifically, in our experiments, the presence of
a non-aggregating strain disrupts a different, normally aggregated, commensal
76
species, lowering its abundance and growth rate and altering its spatial
structure in the intestine. Although the AE-MB4 isolate has reduced spatial
overlap with EN than the parental strain, we find that this does not attenuate
competition, defying the simple expectation that species with different localization
characteristics are likely to have weaker interactions.
The molecular mechanism of the observed disaggregation is currently
unknown. As described in detail elsewhere [163], the Aeromonas-MB4 isolate
investigated here was derived by directed evolution in media rich in N-
Acetylglucosamine (GlcNAc), a sugar prevalent in intestinal mucus [164–167].
Aeromonas is normally highly aggregated in the zebrafish gut, with clusters likely
held together by extracellular polysaccharides and proteins as is typical for biofilms
and other microbial aggregates. It is therefore plausible that in the AE-MB4
isolate, intestinal cues that inhibit production of polysaccharide-digesting enzymes
are ignored, and these enzymes act broadly to degrade the extracellular matrix of
Enterobacter. Testing this hypothesis requires considerable further study and could
benefit from experimental assays such as using microfluidics to identify physical
and chemical conditions in which dissociation occurs and perhaps the role of
geometry of the gut in mediating this competition. Similar phenomena, however,
have been reported in non-intestinal contexts. For example, Streptococcus pyogenes
secretes a protease that disrupts Staphylococcus aureus biofilms [168], the algae C.
reinhardtii can inhibit aggregation of Escherichia coli [66], and mucus enhances in
vitro biofilm formation by Escherichia coli [169] Given the observed co-aggregation
of both Aeromonas and Enterobacter, it is likely that both species share cues and
the disruption of aggregation of one, may directly interfere with aggregation of the
other. Universal cues for inter-species bacterial communication such as AI-2 have
77
been known to promote formation of mixed species biofilms [170, 171]. A study
has found that the suppression of AI-2 expression in one of the species in a dual-
species biofilm results in sparse co-aggregation and decreased biomass [172].
The Aeromonas/Enterobacter interactions we observe in the zebrafish gut
differ strikingly from interactions observed in vitro. In minimal media with 0.4%
GlcNaC, we find co-aggregates of AE-MB4 and EN (Fig 3.2E), with intermingling
of the species at the scale of single cells, a morphology never seen in the gut. The
dissimilarity is not surprising, as both the chemical and physical environment
of the intestine differ from that of a media-filled petri dish, but it highlights the
challenges of inferring in vivo behavior from in vitro assays.
The disaggregation and low abundance induced in Enterobacter by
Aeromonas-MB4 in pairwise competition in the zebrafish gut are attenuated
in the presence of additional bacterial species. The ability of such higher
order interactions to stabilize diverse communities has been noted in many
contexts [70, 71, 73, 159], and the observations described here suggest that
maintenance of spatial structure is a means by which multi-species coexistence
is manifested. It is not currently clear whether this maintenance is facilitated
by the biochemical activity of specific members of the community or by more
complex spatial organization of species that may emerge. Studies involving labeled
subsets of these species, or mutants derived with different aggregation traits, will
illuminate the mechanism.
We also note the role of priority effects in determining community
composition. It is known that early gut microbiome composition in humans
influences whether a new colonizer can establish itself [38]. Studies have linked
a diverse and healthy gut microbiome to colonization resistance [173], for example
78
from invasion by the pathogen Clostridioides difficile [174, 175]. In the larval
zebrafish, we find that in the presence of a diverse pre-existing community,
AE-MB4 interactions with EN and AE are significantly weaker than when all
species are inoculated simultaneously. Imaging reveals the spatial signatures
that distinguish the outcomes of co-inoculation and invasion: AE-MB4 is more
planktonic and less aggregated in the former than the latter, and EN maintains
its aggregated phenotype with little fragmentation in invasion but not in co-
inoculation studies.
The presence of gut bacterial aggregates [51, 89, 91, 92] and their
importance to phenomena such as antibiotic response [53] and resistance to
intestinal contractions [52] have been deduced in past work involving live imaging
in larval zebrafish. This model system enables exploration of the community-level
consequences of altered aggregation of a commensal bacterial species, lending
insight into biophysical and biochemical mechanisms that may be at play more
broadly, for example in the human gut. We suggest that bacterial manipulation
of aggregation state, both their own state and the state of other species, may be a
common feature of gut microbiome dynamics.
Moreover, controlled disruption of bacterial aggregates could have
therapeutic applications. An engineered disruptor, we speculate, could be
introduced to the human gut to displace resident, aggregated microbes, facilitating
their replacement by transplanted or otherwise intentionally provided strains.
Reaching this end will require a more thorough understanding of inter-species
modulation of physical structure.
79
3.4. Materials and Methods
3.4.1. Animal Care
All experiments with zebrafish were performed in accordance with protocols
approved by the University of Oregon Institutional Animal Care and Use
Committee and by following standard protocols [147]
3.4.2. Gnotobiology
Wild-type (AB X TU line) larval zebrafish (Danio rerio) were derived using
protocols described in [47]. The larvae at 0 dpf are washed with antibiotic, bleach
and iodine solutions and then moved to tissue culture flasks containing sterile
embryo medium solution at a density of approximately 1mL/larva. The flasks are
stored in a temperature-controlled room maintained at 28C.
3.4.3. Bacterial strains
Zebrafish gut isolates were previously tagged with green fluorescent
protein (GFP) and dTomato [50]. The various isolates used for this study are
textitAeromonas sp. (ZOR0001), Enterobacter sp. (ZOR0014), Plesiomonas sp.
(ZOR0011) Pseudomonas mendocina (ZWU0006) and Acinetobacter calcoaceticus
(ZOR0008). Aeromonas-MB4 strain (JS168), an isolate of Aeromonas strain
(ZOR0001) was generated through a guided evolution experiment described in
[163]. All bacterial stocks were made in 25% glycerol and maintained at -80◦C.
80
3.4.4. Inoculation of tissue culture flasks
The bacteria from frozen stocks were grown in lysogeny broth (LB medium)
and shaken overnight for approximately 16 hrs in a temperature controlled room at
30◦C. 1mL of the overnight culture was washed twice by centrifuging at 7000g for
2 min and replacing the supernatant with fresh sterile embryo medium each time.
For mono-association and co-inoculation experiments, the species were inoculated
in fish at 5dpf and for the challenge experiments, the first species was inoculated
at 5dpf and the second species is inoculated approximately 24 hrs after, at 6 dpf
such that the flask water concentration of species is approximately 106 CFUs/mL.
3.4.5. Gut dissections and plating
We determined the abundances of bacterial species in the competition
experiments by dissecting and plating the intestine. At 7dpf, zebrafish were
euthanized via hypothermic shock, following which gut dissections were done to
isolate intestinal contents. The intestine was placed in 500µL of sterile embryo
medium and homogenized by blending with 0.5mm zirconium oxide beads.
Dilutions of 10−1 and 10−2 were prepared and 100µ L of these solutions were
spread on tryptic soy agar plates. For experiments with zebrafish inoculated with
> 2 species, guts were plated on Universal HiChrome agar (Sigma Aldrich) to
distinguish species based on a colorimetric indicator. The CFUs on the plates were
counted to ascertain the abundances of the inoculated bacterial species.
3.4.6. Live Imaging using light sheet fluorescence microscopy.
3D imaging of the larval zebrafish gut was done on a home-built setup
described in [89]. The larvae were anesthetized with MS-222 (Syndel) and mounted
81
in 0.6% agarose gel in glass capillaries. Each capillary is inserted into an imaging
chamber containing sterile embryo medium and MS-222 at a concentration of
20µL/mL of embryo medium. Once mounted, the fish were extruded from the
capillaries. Lasers at 488nm and 568nm were used to excite GFP and dTomato
bacteria respectively. Image acquisition involves translating the specimen in z-
steps of 1 micron to capture the entire volume of the zebrafish gut. For time series
experiments, imaging was done overnight at intervals of 20 or 30 minutes. Each
acquired image comprises 2-4 regions that are stitched together using software.
After imaging, the fish were euthanized by placing in an MS-222 solution of
40µL/mL of sterile embryo medium.
3.4.7. In vitro aggregation assay
AE and EN bacterial cultures were grown over night with shaking at
30◦C, normalized to OD600=5.0, then washed 2X in equal volume buffer (sterile
embryo medium). Normalized AE and EN cell suspensions were then back diluted
1:10 into 12-well plates containing buffer supplemented with 0.4% GlcNAc and
incubated for approximately 6 hours with gentle rotation (115 RPM) at 30◦C to
allow co-aggregate formation.
3.4.8. Simulation
The model of bacterial cluster dynamics applied to EN is based on earlier
work investigating bacterial cluster size distributions in the zebrafish gut [92].
We combine four kinetic processes – growth, aggregation, fragmentation, and
expulsion – in a stochastic model simulated with a Gillespie algorithm. Growth
is deterministic and occurs at every time step while aggregation, fragmentation
82
and expulsion are stochastic. Each process is described by a kernel with specific
rate parameters described below.
Growth: We model growth of each cluster as deterministic and following a
simple logistic growth curve, with the total abundance of all clusters barred
from exceeding an overall carrying capacity. The carrying capacity was fixed
at 104 cells, approximately the mono-association abundance of EN. We use
experimentally measured values of the growth rate for both forms of EN, with
rpl = 0.24 hr
−1 for planktonic cells (i.e. clusters of size n = 1) and r −1agg = 0.66 hr
for clusters (n ≥ 2). Growth occurs at every time step with dt being drawn from
an exponential distribution with rate governed by the total overall reaction rate.
Aggregation: Aggregation involves a reaction of two clusters of sizes n and
m combining to form a cluster of size n + m. Aggregation is taken to be size
dependent with the rate of clusters n, m coming together given by:
Kagg = α(n ∗m)1/3 (3.4)
where 1/3 indicates that the collision of clusters is dependent on the cluster
diameter and α = 10−2.5 hr−1. The clusters are randomly grouped in pairs at
every time step and the rate of each aggregation reaction is computed.
Fragmentation: Fragmentation or breakup of a single cell from a cluster of size n
occurs at a size dependent rate-
K 2/3frag = β(n) (3.5)
83
where 2/3 represents the surface area dependence, i.e. cells only on the surface
of clusters can fragment. A value of 0.1 hr−1 corresponds to 10 cells fragmenting
every hour from a cluster of size 1000.
Expulsion: The expulsion process involves removal of clusters from the intestine.
Prior experimental work established that expulsion occurs on average every 10-
12 hours, so the expulsion rate set to λ = 0.1 hr−1agg [51]. The expulsion rate of
λ
non-motile individuals occurs at a rate λ plpl. The expulsion ratio ( ) is variedλagg
λ
logarithmically with pl ∈ [0.25, 100] .
λagg
Mono-association simulation: We first generate the initial mono-association
cluster configuration for EN. Here, we keep β fixed at β = 0.01 hr−1, since in
mono-association we see rare fragmentation of EN and log10(α) = −2.5. The
expulsion rates are λpl = λ
−1
agg = 0.1 hr . Growth of individuals and clusters
occurs at rates rpl = 0.24 hr
−1 and r −1agg = 0.66 hr .
At a given time step, the rates of aggregation, fragmentation and expulsion
for all possible cluster reactions are computed. The total reaction rate i.e. the sum
of all rates determines the jump interval (dt) when the next reaction occurs. At
a given time step, only one cluster reaction is randomly chosen, weighted by the
rate. The chosen reaction is executed, following which (λpldt ) individual cells are
expelled. (Expulsion of individuals is treated deterministically, for computational
speed; note that there are large numbers of individuals, so stochasticity is unlikely
to be important.) Before proceeding to the next time step, all clusters undergo
growth as described previously.
The simulation is carried out until T = 24 hrs, the time at which the EN
population is invaded by the mutant AE-MB4 species. The configurations with
84
large (n > 1000) clusters are saved as the initial configuration for the invasion
simulation.
Invasion simulation: We simulate for different parameters of β ∈ [0.02, 1]
hr−1
λ
in steps of df=0.02 hr−1 and pl ∈ [0.25, 100]. Within a simulation run,
λagg
λ
we choose pl and start with the mono-association cluster size distribution. We
λagg
evolve this distribution until T = 24 hrs (usually when the invasion abundance
data is collected). All the processes i.e. growth, fragmentation, expulsion and
aggregation are simulated exactly as in generating the initial cluster configuration.
The total abundance along with final cluster sizes after 24 hrs are saved. The
mean abundance from 500 fish for the full range of parameters of fragmentation
rate and expulsion ratio are plotted in Fig 3.3D.
3.4.9. Image analysis
Identification of single cells and clusters in 3D images
Identification of single bacterial cells was done using a combination of
machine learning and semi-automated approaches. A convolutional neural network
previously trained on 3D images of single bacteria is used to classify potential
objects as bacteria or noise blobs [109]. The same network architecture was
used and the algorithm was trained on labeled datasets of AE-MB4 and EN cells.
Identification of potential bacteria objects in voxels of size (30x30x10) pixels was
done by using a local maxima finding algorithm in 2D and stitched in 3D. These
potential objects were classified by the network and the classified output was
curated manually.
For aggregate segmentation, the zebrafish gut was first segmented in 2D
using the U-net algorithm which was previously trained on a hand-labeled dataset
85
of zebrafish gut images [110]. After segmentation of the gut, bacterial clusters
within the gut were segmented by applying a background threshold for each image
of the z-stack for segmentation of bacterial aggregates. The total number of cells
in an aggregate is determined by dividing the total fluorescence intensity from the
segmented object by the median intensity of individuals in the image.
Once the net population in single cells and aggregates was ascertained, these
were used to calculate the planktonic fraction, a ratio of the number of individuals
to total abundance. To measure the cumulative cluster size distributions, p(n)
and cluster size histograms depicting probabilities of finding a cell in an n-celled
cluster, we performed a jackknife resampling on the dataset for each experiment
to ascertain the mean and uncertainty in our measurements. The distributions
shown in Fig 3.1 F reflect pooled distributions with all clusters from all fish in an
experiment.
Normalized intensity profiles
After segmentation of the gut using U-net, we defined markers for the
foremost point in the anterior in a 3D image. After applying a single background
threshold for each 3D image, we computed the total intensity in this thresholded
region along the intestinal axis. The normalized intensity profile over several fish is
averaged and shown in Fig 3.1D, 3.5D and 3.6B.
Growth rate measurement
To measure growth rates, we select consecutive 3D scans from fish, wherein
aggregated or planktonic populations of bacteria are undisturbed and there
is no influx or efflux of bacteria from fragmentation or expulsion respectively.
86
We identify single cells and clumps using the image analysis pipeline described
previously. An exponential fit to the abundance of single cells or clusters over time
was used to determine the growth rate of cells and clusters respectively. The mean
and the standard deviation from several fish are reported in Results. .
87
CHAPTER IV
IMMUNE RESPONSE FOR DIFFERENT BACTERIAL SPECIES
4.1. Background
Plenty of evidence exists that the microbial residents within our intestines
interact with the immune system [176–178]. These interactions are crucial for
maintaining intestinal homeostasis. The immune system walks a fine line between
tolerating commensal microbiota and sensing and eliminating pathogens [178].
Which biophysical and biochemical features of bacteria may be interpreted by the
immune system in its decision-making, need further investigation.
There are different types of immune cells that cooperate to defend us from
pathogens. The innate immune system is the body’s fast, nonspecific defense
mechanism comprising an army of different cell types, namely macrophages,
monocytes, dendritic cells, neutrophils, eosinophils. We also have a more specific
and sophisticated immune response triggered by components of the adaptive
immune system, T cells and B cells. We focus here on innate immune responses,
in particular the cytokine TNFα. The adaptive immune response of zebrafish is
not activated until approximately 3 weeks into their development, beyond the age
examined in our studies [179].
Immune cells communicate using compounds called cytokines that may be
pro-inflammatory or anti-inflammatory in nature. Innate immune cells such as
macrophages secrete tumor necrosis factor alpha (TNFα), a proinflammatory
cytokine. This signaling mechanism has been found to play a role in maintaining
88
intestinal homeostasis by stimulating cell-proliferation as well as triggering cell-
death or apoptosis [180].
In previous work specific to zebrafish, immune responses to several different
commensal species have been characterized [90, 101]. In both studies, the highly
motile Vibrio ZWU0020 strain is observed to cause increased inflammation,
showing an influx of neutrophil cells and an almost two orders of magnitude
increased activity of TNFα when compared to germ-free hosts [90, 101].
Interestingly in [90], a bacterial mutant deficient in motility induced a TNFα
response that was comparable to germ-free levels.
These studies of inflammation and bacterial species reveal that the spatial
organization of bacteria impacts their interactions with the immune system. In
this chapter, I focus on characterizing the inflammation response in the form of
TNFα expression and immune-bacteria co-localization in the form of macrophage
localization for a few different commensal bacterial strains. In particular, I focus
on two different bacterial species Aeromonas (AE) and a strain derived from AE
that is resistant to GlcNAc mediated aggregation in the gut, Aeromonas-MB4
(AE-MB4), also used in the studies described in Chapter III. Simultaneous studies
of bacteria and host inflammation in live animals are challenging to perform in
different animal models. The zebrafish model system is well-suited for this aspect
and allows for recording time-varying immune responses with bacterial dynamics.
A specific transgenic line of zebrafish with a fluorescent reporter for TNFα
expression was used in these studies (Tg(tnfa:gfp)) [181]. A different transgenic
line with a fluorescent reporter for macrophages Tg(mpeg1:mCherry) was used to
track macrophages [182]. A cross between these two transgenic lines Tg(tnfa:gfp)
89
x Tg(mpeg1:mCherry) was often used to conduct experiments for concurrently
tracking TNFα expression and macrophage dynamics.
4.2. TNFα response of bacterial strains.
The transgenic fish were derived germ-free as described in Ch. II and III.
Culturing of bacteria and inoculations were also done as described in Ch. II and
II. The strains that were used to assess the TNFα response were Aeromonas
01 (ZOR0001) (AE), Aeromonas-MB4 (AE-MB4), and Enterobacter (ZOR0004)
(EN). Di-associations were also performed with AE and EN, AE-MB4 and
EN. TNFα reporter activity was assessed at 7dpf using light sheet fluorescence
microscopy as described in Ch. I and III. For mono-association experiments all
dTomato strains were used and excited with a 568nm laser, while for the di-
association experiments GFP EN was inoculated along with dTomato AE and
AE-MB4. Exemplar images of the spatial distribution and TNFα expression for
different strains are shown in Fig. 4.1.
To measure the number of TNFα activated cells, I segmented cells from the
images by selecting a user-specified background value for each 3D image. Note
that this approach did lose some dim TNFα cells that are likely in a transitory
state and a more fine-tuned approach would be necessary if those cells needed to
be quantified as well. The numbers of segmented cells for each experiment are
plotted in Fig. 4.1. For fish colonized with AE-MB4, either in mono-association or
di-association with EN, the number of TNFα positive cells is seen to be 3 times
as high as the germ-free condition. In fish colonized with AE and EN strains that
form large aggregates in the mid-gut as described in Ch. III, TNFα activity is
90
FIGURE 4.1. Exemplar maximum intensity projections of TNFα reporter activity
(green) and bacterial spatial distribution (magenta) for (A) AE colonized (B) EN
colonized (C) AE-MB4 colonized (D) germ-free tnfa:gfp transgenic fish. (E) and
(F) are maximum intensity projections from fish co-inoculated with AE(magenta)
EN (green) and AE-MB4(magenta) and EN (green). The TNFα reporter activity
and macrophages (only E) are shown in green and magenta respectively. Scale
bar: 50µ m (E) The total number of TNFα+ cells in germ-free fish, fish mono-
associated with AE, AE-MB4, EN, co-colonized by EN and AE-MB4, EN and AE.
N (from left to right) =[10, 6, 17,7,12, 10] fish.
91
comparable to germ-free fish, suggesting these strains do not produce a strong pro-
inflammatory response.
Interestingly, we also found that the TNFα response in the AE-EN
diassociation was also around 2-3 times higher compared to fish mono-colonized
with AE or EN alone.
Although there seems to be a high degree of variability in TNFα response
for fish colonized with AE-MB4, there are clear patterns of greater inflammation
compared to the response in fish with AE and EN.The differences in spatial
organization of AE and AE-MB4 characterized in Ch II could play a role in AE-
MB4 inflammation as reported with Vibrio species described in [90]. AE-MB4
is deficient in its mucus sensing pathway that results in its altered aggregation
behavior in the gut, which may directly or indirectly be responsible for increased
inflammation. More studies investigating AE-MB4 will illuminate the specific
conditions wherein this bacterial strain can trigger increased inflammation.
4.3. Macrophage bacteria interactions
Using the double transgenic- Tg(tnfa:gfp) x Tg(mpeg1:mCherry), I also
observed specific macrophage localization patterns in fish mono-associated with
AE-MB4 or in di-association with AE-MB4 and EN. We observed several examples
of ‘macrophage swarms’ in the intestinal bulb shown in Fig. 4.2. These swarms are
oftentimes localized adjacent to regions where AE-MB4 cells are localized. In rare
instances, these swarms show TNFα activity [Fig. 4.2]. Many questions regarding
what feature of AE-MB4 triggers such swarm formation and what function
these swarms serve, remain to be answered. Recent work by others has explored
swarms of innate immune cells, for example neutrophil swarms suppressing fungal
92
growth [Hopke et al. 2020]. Studies in zebrafish have reported the formation
of granulomas or well-organized macrophage aggregates upon infection with
Schistosoma eggs [183, 184]. Both the TNFα and macrophage observations open
up many more possibilities for deeper exploration.
FIGURE 4.2. Examples of macrophage swarms (large clumps in magenta) from
maximum intensity projections of the gut of different fish that were initially
inoculated with EN and then challenged with AE-MB4 after 24 hrs. The gut
outline is shown in dashed white. A, E and F are from the the midgut regions,
while B, C and D are from the foregut region of the imaged fish. The region in F
shows TNFα+ cells within the macrophage swarm. Scale bar: 50µm
93
4.4. Conclusion
Visualizing and quantifying immune system dynamics concurrently with
bacteria can yield many useful insights into host-microbe interactions. I show that
the planktonic species AE-MB4 stimulates increased inflammation compared to
its aggregated parental strain AE by measuring TNFα activity, suggesting that
the single trait of deficient mucus sensing and all its downstream effects can be
sufficient to alter host immune response in the vertebrate intestine. Other recent
work submitted for review also corroborates the increased inflammation seen
with the AE-MB4 strain [163]. Earlier investigations of inflammation with motile
Vibrio species showing similar planktonic behavior and localization in the anterior
region of the intestine may suggest that generic features of spatial structure and
localization of species are determinants of host immune response [90, 101].
Patterns in macrophage spatial dynamics show clustering of macrophages
upon colonization by AE-MB4. In some instances, these clusters are TNFα+
as well. It would be useful to visualize the dynamics of swarm formation and
dissociation in conjunction with bacterial dynamics. Some recent studies
characterize neutrophil migration behaviors in zebrafish, noting pioneer neutrophils
in swarms [185]. Further, what timescales dictate the formation of these swarms
are indicative of the response times of such dendritic cells and other aspects such
as the diffusion of cytokines.
The TNFα response with AE-MB4 translates into di-association experiments
with EN, retaining its proinflammatory capacity with this species. It would be
interesting to study how previously discovered anti-inflammatory zebrafish isolates
such as Shewenella [101] interact with AE-MB4 and whether they can suppress
94
TNFα response. Such studies could help gain insight into whether adding select
microbial members can restore TNFα activity to normal levels.
It’s unclear why the combination of AE and EN also produced increased
inflammation, something not observed in AE and EN mono-colonized fish. What
aspects of competition and whether spatial structure changes or altered signaling
cues may be contributing to such responses requires a characterization of this
competition and spatial structure. These immune cell investigations pave the way
for more such studies visualizing and characterizing host immune response in living
organisms.
95
CHAPTER V
CONCLUSIONS AND FUTURE WORK
Combining 3D live imaging, zebrafish transgenics, and the genetic tools
developed for commensal bacterial strains has enabled addressing questions
regarding interaction dynamics, physical structure, and immune activation that
would be impossible to study in humans and other model organisms. There are
many directions for future studies that build on insights from this work.
Through mapping interactions between a consortium of five native gut-
bacterial isolates, I found the existence of higher order interactions that weaken
strong competitive interactions and promote coexistence in the zebrafish gut. A
question that arises from this work is how these results translate to communities
of a greater number of species or different sets of species. These questions would
require more thoughtful experimentation, however the fact that these 4-5 species
communities are sufficient to see differences in interaction patterns is telling of the
merits of performing such controlled, tractable studies.
The similarity in abundance distributions of both four and five species
experiments as well as across different species in these experiments is striking
and indicative of species-specific features becoming less apparent in diverse
communities. In addition to this, weak competition in five-species communities
suggests a shift towards a more neutral community, or a community whose
composition is primarily determined by stochastic ecological processes. Studies to
specifically uncover the role of these stochastic processes, such as host colonization
and transmission between hosts will illuminate to what extent these processes
govern composition in complex communities.
96
One may also ask, whether the addition of a strong competitor can break
patterns of coexistence in this model community. Using the interactions mapped in
Chapter II, we discovered specific pairs such as Aeromonas-Enterobacter showing
strong competition. This pair, in combination with an Aeromonas isolate (AE-
MB4) evolved by colleagues Jarrod Smith and Karen Guillemin that lacks the
trait of aggregating in a sugar found in intestinal mucus, was used to test how
altered aggregation impacts community structure. Through this study, we found
a novel interaction mechanism, namely, induced dissociation of Enterobacter
aggregates that are prone to intestinal expulsion. The study underscores the
ability of a single species to disrupt community structure and for communities
to restore themselves in the presence of diverse members. More exploration of how
these diverse communities mitigate the dramatic impact AE-MB4 has on two-
species communities would be telling of the possible mechanisms of interactions
underlying the human gut microbiome. My colleague Jarrod Smith has engineered
a genetic switch that converts the bacterial behavior of wild-type Aeromonas from
aggregated to planktonic. Using powerful genetic tools such as these, one can
perturb microbial communities in live fish and visualize community dynamics, an
impossible pursuit in humans or other model systems.
Imaging the spatial structure of multi-species communities has shed insights
into different mechanisms of co-aggregation of bacterial species in vivo vs in vitro.
Along with Jarrod, I observed differences in the structure of bacterial aggregates
of AE and EN in vivo and in vitro. For instance, in vivo, these aggregates were
often conglomerates of smaller single-species clumps rather than homogeneously
distributed cells of both species, suggesting intestinal flow dynamics, especially
mixing of isolated clumps is an important factor in co-aggregate formation.
97
In contrast, imaging in vitro aggregates show that the cells of these individual
species are distributed more or less uniformly across the aggregate. More thorough
comparative studies of cluster dynamics in vitro and in vivo can illuminate the role
of the host in facilitating bacterial aggregation.
Along with the characterization of cluster structure, an understanding of
the constituents of these bacterial clusters would reveal the role of possible host-
derived structural components such as mucus in aggregation processes. During
some recent in vivo imaging studies of live fish stained with WGA (wheat germ
agglutinin), a protein that binds to particular sugars, some of which are found in
mucus, Jarrod and I found that bacterial aggregates are often times correlated
with regions found to be rich in mucus sugars. Such staining experiments can
provide details into the composition of these aggregates and perhaps what host
cues drive aggregation.
Finally, designing studies to map host immune responses of various members
and their aggregation properties illuminate what physical and biochemical factors
of microbes are being interpreted by the immune system. Through studies
comparing inflammation in fish colonized by AE and AE-MB4, clear differences
in TNFα response are observed, similar to the observations in [90]. Such
experiments outline the possibility of general spatial structure determinants of
immune response. More studies focussing on different species, their inflammation
dynamics and macrophage activity would help decipher species-specific and generic
characteristics of host-microbe interactions.
Pro-inflammatory bacterial strains such AE-MB4 can be used to address
many other questions in microbiome research. Previous studies to study the effects
of antibiotic exposure on the microbiota discovered that motile bacteria undergo a
98
stress response and clump in response to antibiotics [53]. Antibiotics have emerged
as a powerful tool to perturb intestinal communities. How antibiotics differentially
impact species and can alter species interactions and intestinal community
composition remains to be known. Further, studies have shown that after a
treatment of tetracycline, zebrafish show suppressed production of cytokines [186].
Investigations of how the immune system responds to pro-inflammatory species
such as Vibrio and AE-MB4 after antibiotic exposure is unknown, a path that
could yield lots of novel information about antibiotic exposure and the immune
system.
The combination of using zebrafish as a model system, the exciting avenue
of gut microbiome research, novel bacterial engineering approaches, imaging and
quantitative modeling leads to an unending set of possibilities for future studies.
Specifically in the area of biophysics, the intersection of all of these approaches has
the potential to reveal general rules underlying these complex biological systems
that can be of great significance to human health. With increasing interest in gut
microbiome research, the need for more quantitative approaches and models by
physicists to characterize and simplify the complexity in these systems is certain
and will only spur more discoveries about the mysteries of intestinal communities.
99
APPENDIX
Species j
FIGURE A.1. The total bacterial load for different di-association experiments.The
total bacterial load for different di-association experiments, expressed as the mean
and standard deviation of log10 (total CFUs). The values on the diagonal are the
mean load from mono-association experiments.
100
Species i
A B
Species j Species j
FIGURE A.2. Pairwise interaction coefficients for two-species and five-species
experiments using the log-transformed model. The mean pairwise interaction
coefficients CIIij (A) and C
V
ij (B) showing the effect of species j on species i
calculated using the log-transformed abundance model. The standard deviations
are calculated using a subsampling approach described in Methods.
Species j
FIGURE A.3. Pairwise interaction coefficients from in vitro two-species
experiments calculated using the log-transformed abundance model. The matrix
of interaction coefficients showing the mean and standard deviation of CIIij from in
vitro competition experiments.
101
Species i
Species i
Species i
α = 0.1 α = 1.5
60
40
10000 20
0 0
-20
-10000 -40
-60
α = 0.5 α = 2
2000 10
1000 5
0 0
-1000 -5
-2000 -10
α = 1
200
0
-200
Species j 
FIGURE A.4. Pairwise interaction coefficients for select α values for two and five-
species experiments using the power law model. Interaction coefficients generated
from the linear absolute abundance model for are compared for the one-to-two
species (left) and four-to-five species (right) experiments. The legend at the
bottom-right shows the species labels for the rows and columns.
102
Species i 
Species j Species j
FIGURE A.5. Pairwise interaction coefficients for two and five-species experiments
using a linear model. The mean and standard deviation of the interaction
coefficients for the two species CIIij and five species C
V
ij experiments calculated
using a model linear in absolute species abundance.
103
Species i
Species i
A B
Predicted Frequency
Species
FIGURE A.6. Predicted five species distributions from the linear model. A.
Predicted (blue xs) and measured (brown circles) absolute abundance distributions
for five-species inoculation experiments calculated from the linear model for
pairwise interactions, for each of the five species. The means of the predicted
and measured distributions excluding nulls are shown using bold blue circles
and brown square markers, respectively, with error bars indicating the standard
deviation. The dotted line indicates the experimental detection limit of 25 cells.
The predicted distributions are generated from sampling the interaction coefficient
distributions as described in Methods, while the experimental distributions
comprise abundances from N = 202 fish. B. The observed occurrence frequencies
of each species in five-species experiments plotted against the predicted frequencies
generated from the linear model.
104
CFUs/gut+1
 Measured Frequency
AE AE-MB4
FIGURE A.7. In vitro biofilm assay for AE (top panel) and AE-MB4 (bottom
panel) in 0.4% GlcNAc solution. AE forms macroscopic aggregates in 0.4%
GlcNAc (top-right) while AE-MB4 (bottom-right) is unable to form biofilms in
the same medium.
105
Cluster size(n)
FIGURE A.8. The cumulative cluster size distributions (p(cluster size ≥ n)) for
(i) AE-MB4 (ii) EN (iii) AE and (iv) all three species in mono-association. The
magenta, green and orange curves in (i), (ii) and (iii) are distributions calculated
from clusters found in single fish in each mono-association experiment. The black
curve in (i)-(iii) shows the pooled distribution calculated from clusters found in
all fish. (iv) shows the pooled distribution for each of the species with all three
showing power law behavior with slope m = 0.8 ± 0.1 for AE-MB4 and AE and m
= 0.9 ± 0.2 for EN in the regime n < 102. Data are from N = 6, 7 and 8 fish for
AE, EN and AE-MB4, respectively.
106
p(cluster size ≥ n)
Cluster size(n)
FIGURE A.9. The probability of finding a cell in an n-cell cluster for AE (orange),
EN (green) and AE-MB4 (magenta) in mono-association. Circles and error bars
indicate the mean and uncertainties, calculated using jack-knife resampling. Tick
marks indicate bin intervals, e.g. the three bars between 0 and 100 correspond to
EN, AE and AE-MB4 probabilities to be in clusters of size n = (0-1]. Data are
from N = 6, 7 and 8 fish for AE, EN and AE-MB4, respectively.
FIGURE A.10. Maximum intensity projection of 3D image of a larval zebrafish
gut colonized with EN (green) at 5dpf and challenged with AE (orange) at 6dpf.
Overlapping populations are in yellow. Bar:100µm
107
Probability of a cell in
 a cluster of size n
REFERENCES CITED
[1] R. W. J. Cavicchioli, Ricardo et al. Scientists’ warning to humanity:
microorganisms and climate change. Nature Reviews Microbiology, 17(9),
September 2019.
[2] J. U. Scher and S. B. Abramson. The microbiome and rheumatoid arthritis.
Nature Reviews Rheumatology, 7(10), October 2011.
[3] A. B. Shreiner et al. The gut microbiome in health and in disease. Current
opinion in gastroenterology, 31(1):69–75, January 2015.
[4] K. J. Pflughoeft and J. Versalovic. Human Microbiome in Health and Disease.
Annual Review of Pathology: Mechanisms of Disease, 7(1):99–122, 2012.
[5] R. D. Moloney et al. The microbiome: stress, health and disease. Mammalian
Genome: Official Journal of the International Mammalian Genome Society,
25(1-2):49–74, February 2014.
[6] W. Zhu et al. Precision editing of the gut microbiota ameliorates colitis. Nature,
553(7687):208–211, January 2018.
[7] B. Wang et al. The Human Microbiota in Health and Disease. Engineering,
3(1):71–82, February 2017.
[8] J. Durack and S. V. Lynch. The gut microbiome: Relationships with disease and
opportunities for therapy. Journal of Experimental Medicine, 216(1), January
2019.
[9] M. Rajilić, Stojanović et al. Global and Deep Molecular Analysis of Microbiota
Signatures in Fecal Samples From Patients With Irritable Bowel Syndrome.
Gastroenterology, 141(5):1792–1801, November 2011.
[10] K. Honda and D. R. Littman. The microbiome in infectious disease and
inflammation. Annual Review of Immunology, 30:759–795, 2012.
[11] C. L. Boulangé et al. Impact of the gut microbiota on inflammation, obesity,
and metabolic disease. Genome Medicine, 8:42, April 2016.
[12] J. H. Hill et al. A conserved bacterial protein induces pancreatic beta cell
expansion during zebrafish development. eLife, 5, December 2016.
[13] I. Semova et al. Microbiota regulate intestinal absorption and metabolism of
fatty acids in the zebrafish. Cell Host & Microbe, 12(3):277–288, September
2012.
108
[14] K. Martinez-Guryn et al. Small Intestine Microbiota Regulate Host Digestive
and Absorptive Adaptive Responses to Dietary Lipids. Cell Host & Microbe,
23(4):458–469.e5, 2018.
[15] H. Sokol et al. Faecalibacterium prausnitzii is an anti-inflammatory commensal
bacterium identified by gut microbiota analysis of Crohn disease patients.
Proceedings of the National Academy of Sciences of the United States of
America, 105(43):16731–16736, October 2008.
[16] N. Arpaia et al. Metabolites produced by commensal bacteria promote
peripheral regulatory T-cell generation. Nature, 504(7480):451–455,
December 2013.
[17] N. Kamada and G. Nunez. Regulation of the Immune System by the Resident
Intestinal Bacteria. Gastroenterology, 146(6):1477–1488, May 2014.
[18] S. K. Mazmanian et al. An Immunomodulatory Molecule of Symbiotic Bacteria
Directs Maturation of the Host Immune System. Cell, 122(1):107–118, July
2005.
[19] L. Ye et al. High fat diet induces microbiota-dependent silencing of
enteroendocrine cells. eLife, 8, December 2019.
[20] J. Yan et al. Gut microbiota induce IGF-1 and promote bone formation and
growth. Proceedings of the National Academy of Sciences,
113(47):E7554–E7563, November 2016.
[21] C. L. Maynard et al. Reciprocal Interactions of the Intestinal Microbiota and
Immune System. Nature, 489(7415):231–241, September 2012.
[22] A. Everard et al. Cross-talk between Akkermansia muciniphila and intestinal
epithelium controls diet-induced obesity. Proceedings of the National
Academy of Sciences of the United States of America, 110(22):9066–9071,
May 2013.
[23] T. C. Scharschmidt et al. A wave of regulatory T cells into neonatal skin
mediates tolerance to commensal microbes. Immunity, 43(5):1011–1021,
November 2015.
[24] D. Perlman et al. Concepts and Consequences of a Core Gut Microbiota for
Animal Growth and Development. Annual Review of Animal Biosciences,
10(1), 2022.
[25] D. Haller, editor. The Gut Microbiome in Health and Disease. Springer
International Publishing, Cham, 2018.
109
[26] K. D. Kohl and H. V. Carey. A place for hostâmicrobe symbiosis in the
comparative physiologist’s toolbox. Journal of Experimental Biology, 219(22),
November 2016.
[27] C. P. Tamboli et al. Dysbiosis in inflammatory bowel disease. Gut, 53(1):1–4,
January 2004.
[28] D. I. Tedjo et al. The fecal microbiota as a biomarker for disease activity in
Crohnâs disease. Scientific Reports, 6(1):35216, October 2016.
[29] N. A. Nagalingam and S. V. Lynch. Role of the Microbiota in Inflammatory
Bowel Diseases. Inflammatory Bowel Diseases, 18(5):968–984, May 2012.
[30] M. Saleh and C. O. Elson. Experimental Inflammatory Bowel Disease: Insights
into the Host-Microbiota Dialog. Immunity, 34(3):293–302, March 2011.
[31] C. Abraham and R. Medzhitov. Interactions Between the Host Innate Immune
System and Microbes in Inflammatory Bowel Disease. Gastroenterology,
140(6):1729–1737, 2011.
[32] J. Cremer et al. Effect of flow and peristaltic mixing on bacterial growth in a
gut-like channel. Proceedings of the National Academy of Sciences, 113(41),
October 2016.
[33] M. Arnoldini et al. Bacterial growth, flow, and mixing shape human gut
microbiota density and composition. Gut Microbes, 9(6), May 2018.
[34] F. Vuik et al. Composition of the mucosa-associated microbiota along the
entire gastrointestinal tract of human individuals. United European
Gastroenterology Journal, 7(7):897–907, 2019.
[35] J. Lloyd-Price et al. The healthy human microbiome. Genome Medicine, 8:51,
April 2016.
[36] A. Almeida et al. A new genomic blueprint of the human gut microbiota.
Nature, 568(7753):499–504, 2019.
[37] C. Milani et al. The First Microbial Colonizers of the Human Gut:
Composition, Activities, and Health Implications of the Infant Gut
Microbiota. Microbiology and molecular biology reviews: MMBR,
81(4):e00036–17, December 2017.
[38] D. Sprockett et al. Role of priority effects in the early-life assembly of the gut
microbiota. Nature Reviews Gastroenterology & Hepatology, 15(4), April
2018.
110
[39] J. L. Sonnenburg and F. BÀckhed. Dietâmicrobiota interactions as
moderators of human metabolism. Nature, 535(7610), July 2016.
[40] L. A. David et al. Diet rapidly and reproducibly alters the human gut
microbiome. Nature, 505, January 2014.
[41] C. De Filippo et al. Impact of diet in shaping gut microbiota revealed by a
comparative study in children from Europe and rural Africa. Proceedings of
the National Academy of Science, 107, August 2010.
[42] G. D. Wu et al. Linking Long-Term Dietary Patterns with Gut Microbial
Enterotypes. Science, 334, October 2011.
[43] K. A. Dill-McFarland et al. Close social relationships correlate with human gut
microbiota composition. Scientific Reports, 9(1), January 2019.
[44] P. W. O’Toole and I. B. Jeffery. Gut microbiota and aging. Science (New York,
N.Y.), 350(6265):1214–1215, December 2015.
[45] S. J. Jackson et al. Does age matter? The impact of rodent age on study
outcomes. Laboratory Animals, 51(2).
[46] A. Schlegel and D. Y. R. Stainier. Lessons from âLowerâ Organisms: What
Worms, Flies, and Zebrafish Can Teach Us about Human Energy
Metabolism. PLOS Genetics, 3(11).
[47] E. Melancon et al. Best practices for germ-free derivation and gnotobiotic
zebrafish husbandry. Methods in cell biology, 138:61–100, 2017.
[48] K. Dooley and L. I. Zon. Zebrafish: a model system for the study of human
disease. Current Opinion in Genetics & Development, 10(3):252–256, June
2000.
[49] J. F. Rawls et al. Gnotobiotic zebrafish reveal evolutionarily conserved
responses to the gut microbiota. Proceedings of the National Academy of
Sciences, 101(13), March 2004.
[50] T. J. Wiles et al. Modernized Tools for Streamlined Genetic Manipulation and
Comparative Study of Wild and Diverse Proteobacterial Lineages. mBio,
9(5), November 2018.
[51] T. J. Wiles et al. Host Gut Motility Promotes Competitive Exclusion within a
Model Intestinal Microbiota. PLOS Biology, 14(7):e1002517, July 2016.
[52] S. L. Logan et al. The Vibrio cholerae type VI secretion system can modulate
host intestinal mechanics to displace gut bacterial symbionts. Proceedings of
the National Academy of Sciences, 115(16):E3779, April 2018.
111
[53] B. H. Schlomann et al. Sublethal antibiotics collapse gut bacterial populations
by enhancing aggregation and expulsion. Proceedings of the National
Academy of Sciences, 116(43):21392, October 2019.
[54] C. Darwin. The Origin of Species. P. F. Collier & son, 1909.
[55] P. J. Morin. Community Ecology. Wiley-Blackwell, Chichester, West Sussex ;
Hoboken, NJ, 2nd edition edition, August 2011.
[56] F. E. Clements. Plant succession; an analysis of the development of vegetation,.
Carnegie Institution of Washington,, Washington,, 1916.
[57] C. S. C. S. Elton. The pattern of animal communities. Methuen ;, London,
1966.
[58] R. T. Paine. Food Webs: Linkage, Interaction Strength and Community
Infrastructure. Journal of Animal Ecology, 49(3):667–685, 1980.
[59] R. T. Paine. Food Web Complexity and Species Diversity. The American
Naturalist, 100(910):65–75, 1966.
[60] R. T. Paine. Energy Flow in a Natural Population of the Herbivorous
Gastropod Tegula Funerralis1. Limnology and Oceanography, 16(1), 1971.
[61] R. T. Paine. Intertidal community structure : Experimental studies on the
relationship between a dominant competitor and its principal predator.
Oecologia, 15(2):93–120, June 1974.
[62] C. K. Fisher and P. Mehta. Identifying Keystone Species in the Human Gut
Microbiome from Metagenomic Timeseries using Sparse Linear Regression.
PLoS ONE, 9(7), July 2014.
[63] G. G. Mittelbach and B. J. McGill. Community Ecology. Oxford University
Press, May 2019.
[64] J. H. Vandermeer. The Competitive Structure of Communities: An
Experimental Approach with Protozoa. Ecology, 50(3):362–371, 1969.
[65] J. Friedman et al. Community structure follows simple assembly rules in
microbial microcosms. Nature Ecology & Evolution, 1(5):1–7, March 2017.
[66] H. Mickalide and S. Kuehn. Higher-Order Interaction between Species Inhibits
Bacterial Invasion of a Phototroph-Predator Microbial Community. Cell
Systems, 9(6):521–533.e10, December 2019.
[67] M. Morin et al. Changes in the genetic requirements for microbial interactions
with increasing community complexity. eLife, 7:e37072, September 2018.
112
[68] A. O. Lopez et al. Interspecies bacterial competition determines community
assembly in the c. elegans intestine. bioRxiv, pp. 535633, January 2019.
[69] A. L. Gould et al. Microbiome interactions shape host fitness. Proceedings of
the National Academy of Sciences, 115(51):E11951–E11960, December 2018.
[70] E. Bairey et al. High-order species interactions shape ecosystem diversity.
Nature Communications, 7(1):1–7, August 2016.
[71] J. M. Levine et al. Beyond pairwise mechanisms of species coexistence in
complex communities. Nature, 546(7656):56–64, June 2017.
[72] J. Grilli et al. Higher-order interactions stabilize dynamics in competitive
network models. Nature, 548(7666):210–213, August 2017.
[73] A. Posfai et al. Metabolic Trade-Offs Promote Diversity in a Model Ecosystem.
Physical Review Letters, 118(2):028103, January 2017.
[74] C. A. Lozupone et al. Diversity, stability and resilience of the human gut
microbiota. Nature, 489(7415), September 2012.
[75] P. Mehta et al. A high-bias, low-variance introduction to machine learning for
physicists. Physics reports, 810, 2019.
[76] W. R. Harcombe et al. Metabolic Resource Allocation in Individual Microbes
Determines Ecosystem Interactions and Spatial Dynamics. Cell Reports,
7(4):1104–1115, May 2014.
[77] T. Taillefumier et al. Microbial consortia at steady supply. eLife, 6, May 2017.
[78] B. G. Weiner et al. Spatial ecology of territorial populations. Proceedings of the
National Academy of Sciences, 116(36), September 2019.
[79] H. J. Kim et al. Defined spatial structure stabilizes a synthetic multispecies
bacterial community. Proceedings of the National Academy of Sciences,
105(47), November 2008.
[80] V. Torsvik et al. High diversity in DNA of soil bacteria. Applied and
Environmental Microbiology, 56(3):782–787, March 1990.
[81] D. E. Dykhuizen. Santa Rosalia revisited: Why are there so many species of
bacteria? Antonie van Leeuwenhoek, 73(1):25–33, January 1998.
[82] C. Tropini et al. The Gut Microbiome: Connecting Spatial Organization to
Function. Cell Host & Microbe, 21(4), April 2017.
[83] J. Nguyen et al. Cause or effect? The spatial organization of pathogens and the
gut microbiota in disease. Microbes and Infection, 23(6-7), August 2021.
113
[84] G. P. Donaldson et al. Gut biogeography of the bacterial microbiota. Nature
Reviews. Microbiology, 14(1):20–32, January 2016.
[85] T. C. Fung et al. Anatomical localization of commensal bacteria in immune cell
homeostasis and disease. Immunological Reviews, 260(1), 2014.
[86] T. Bjarnsholt. The role of bacterial biofilms in chronic infections. APMIS,
121(s136), 2013.
[87] T. Bjarnsholt et al. Pseudomonas aeruginosa tolerance to tobramycin,
hydrogen peroxide and polymorphonuclear leukocytes is quorum-sensing
dependent. Microbiology (Reading, England), 151(Pt 2), February 2005.
[88] K. M. Ottemann and J. F. Miller. Roles for motility in bacteria host
interactions. Molecular Microbiology, 24(6), 1997.
[89] M. Jemielita et al. Spatial and temporal features of the growth of a bacterial
species colonizing the zebrafish gut. MBio, 5(6):e01751–14, 2014.
[90] T. J. Wiles et al. Swimming motility of a gut bacterial symbiont promotes
resistance to intestinal expulsion and enhances inflammation. PLOS Biology,
18(3), March 2020.
[91] B. H. Schlomann et al. Bacterial Cohesion Predicts Spatial Distribution in the
Larval Zebrafish Intestine. Biophysical Journal, 115(11):2271–2277, December
2018.
[92] B. H. Schlomann and R. Parthasarathy. Gut bacterial aggregates as living gels.
eLife, 10:e71105, September 2021.
[93] N. T. Porter and E. C. Martens. The Critical Roles of Polysaccharides in Gut
Microbial Ecology and Physiology. Annual Review of Microbiology,
71:349–369, September 2017.
[94] A. J. Silva et al. Haemagglutinin/protease expression and mucin gel
penetration in El Tor biotype Vibrio cholerae. Microbiology (Reading,
England), 149(Pt 7):1883–1891, 2003.
[95] A. Marcobal et al. A refined palate: Bacterial consumption of host glycans in
the gut. Glycobiology, 23(9), September 2013.
[96] L. Wrzosek et al. Bacteroides thetaiotaomicron and Faecalibacterium
prausnitzii influence the production of mucus glycans and the development of
goblet cells in the colonic epithelium of a gnotobiotic model rodent. BMC
biology, 11:61, May 2013.
114
[97] T. Rossy et al. Pseudomonas aeruginosa contracts mucus to rapidly form
biofilms in tissue-engineered human airways. May 2022.
[98] R. M. Landry et al. MucinâPseudomonas aeruginosa interactions promote
biofilm formation and antibiotic resistance. Molecular Microbiology,
59(1):142–151, 2006.
[99] S. Pirr and D. Viemann. Host Factors of Favorable Intestinal Microbial
Colonization. Frontiers in Immunology, 11, 2020.
[100] K. Stagaman et al. The role of adaptive immunity as an ecological filter on
the gut microbiota in zebrafish. The ISME Journal, 11(7):1630–1639, 2017.
[101] A. S. Rolig et al. Individual members of the microbiota disproportionately
modulate host innate immune responses. Cell host & microbe, 18(5):613–620,
November 2015.
[102] M. Jemielita et al. Comparing phototoxicity during the development of a
zebrafish craniofacial bone using confocal and light sheet fluorescence
microscopy techniques. Journal of biophotonics, 6(0):10.1002/jbio.201200144,
December 2013.
[103] R. M. Power and J. Huisken. A guide to light-sheet fluorescence microscopy
for multiscale imaging. Nature Methods, 14(4):360–373, March 2017.
[104] P. J. Keller et al. Reconstruction of zebrafish early embryonic development by
scanned light sheet microscopy. Science (New York, N.Y.),
322(5904):1065–1069, November 2008.
[105] P. J. Keller and H.-U. Dodt. Light sheet microscopy of living or cleared
specimens. Current Opinion in Neurobiology, 22(1):138–143, February 2012.
[106] R. Parthasarathy. Monitoring microbial communities using light sheet
fluorescence microscopy. Current Opinion in Microbiology, 43:31–37, June
2018.
[107] Y. Wan et al. Light-Sheet Microscopy and Its Potential for Understanding
Developmental Processes. Annual Review of Cell and Developmental Biology,
35(1), 2019.
[108] Y. Gong et al. A fully water coupled oblique light-sheet microscope. Scientific
Reports, 12(1), April 2022.
[109] E. A. Hay and R. Parthasarathy. Performance of convolutional neural
networks for identification of bacteria in 3D microscopy datasets. PLOS
Computational Biology, 14(12), December 2018.
115
[110] O. Ronneberger et al. U-Net: Convolutional Networks for Biomedical Image
Segmentation. arXiv:1505.04597 [cs], May 2015.
[111] E. A. Hay. Identifying Gut Bacteria and Their Interactions Using Deep
Learning Based Image Analysis and Gnotobiotic Experiments. Ph.D.,
University of Oregon, Ann Arbor, 2019.
[112] L. Dethlefsen and D. A. Relman. Incomplete recovery and individualized
responses of the human distal gut microbiota to repeated antibiotic
perturbation. Proceedings of the National Academy of Sciences,
108(Supplement 1), March 2011.
[113] S. R. Modi et al. Antibiotics and the gut microbiota. The Journal of Clinical
Investigation, 124(10):4212–4218, October 2014.
[114] L. Zhao et al. Gut bacteria selectively promoted by dietary fibers alleviate
type 2 diabetes. Science, 359(6380):1151, March 2018.
[115] R. N. Carmody et al. Diet dominates host genotype in shaping the murine
gut microbiota. Cell host & microbe, 17(1):72–84, January 2015.
[116] Y. Shang et al. Inferring interactions in complex microbial communities from
nucleotide sequence data and environmental parameters. PloS One,
12(3):e0173765, 2017.
[117] A. Carr et al. Use and abuse of correlation analyses in microbial ecology. The
ISME Journal, 13(11):2647–2655, November 2019.
[118] D. Lovell et al. Proportionality: A Valid Alternative to Correlation for
Relative Data. PLOS Computational Biology, 11(3), March 2015.
[119] N. Zmora et al. Personalized Gut Mucosal Colonization Resistance to Empiric
Probiotics Is Associated with Unique Host and Microbiome Features. Cell,
174(6):1388–1405.e21, 2018.
[120] A. Aranda-DÃaz et al. Bacterial interspecies interactions modulate
pH-mediated antibiotic tolerance. eLife, 9, January 2020.
[121] A. R. Burns et al. Contribution of neutral processes to the assembly of gut
microbial communities in the zebrafish over host development. The ISME
journal, 10(3):655–664, March 2016.
[122] E. M. Flores et al. The zebrafish as a model for gastrointestinal tractâmicrobe
interactions. Cellular Microbiology, 22(3), 2020.
[123] J. T. Wootton. The Nature and Consequences of Indirect Effects in Ecological
Communities. Annual Review of Ecology and Systematics, 25:443–466, 1994.
116
[124] I. Billick and T. J. Case. Higher Order Interactions in Ecological
Communities: What Are They and How Can They be Detected? Ecology,
75(6):1530–1543, 1994.
[125] A. Sanchez. Defining Higher-Order Interactions in Synthetic Ecology: Lessons
from Physics and Quantitative Genetics. Cell Systems, 9(6):519–520, 2019.
[126] V. C. Ozgen et al. Spatial interference scale as a determinant of microbial
range expansion. Science Advances, 4(11):eaau0695, November 2018.
[127] R. M. Stubbendieck and P. D. Straight. Multifaceted Interfaces of Bacterial
Competition. Journal of Bacteriology, 198(16):2145–2155, August 2016.
[128] A. Prindle et al. Ion channels enable electrical communication within bacterial
communities. Nature, 527(7576):59–63, November 2015.
[129] C. Hoffmann et al. Archaea and Fungi of the Human Gut Microbiome:
Correlations with Diet and Bacterial Residents. PLoS ONE, 8(6), June 2013.
[130] J. A. Steele et al. Marine bacterial, archaeal and protistan association
networks reveal ecological linkages. The ISME Journal, 5(9), September
2011.
[131] B. H. Schlomann. Stationary moments, diffusion limits, and extinction times
for logistic growth with random catastrophes. Journal of Theoretical Biology,
454:154–163, 2018.
[132] R. R. Stein et al. Ecological Modeling from Time-Series Inference: Insight
into Dynamics and Stability of Intestinal Microbiota. PLOS Computational
Biology, 9(12), December 2013.
[133] S. Marino et al. Mathematical modeling of primary succession of murine
intestinal microbiota. Proceedings of the National Academy of Sciences,
111(1), January 2014.
[134] A. Sanchez-Gorostiaga et al. High-order interactions distort the functional
landscape of microbial consortia. PLOS Biology, 17(12), December 2019.
[135] L. R. Lawlor. Direct and indirect effects of n-species competition. Oecologia,
43(3):355–364, December 1979.
[136] M. Dungan. Three-Way Interactions: Barnacles, Limpets, and Algae in a
Sonoran Desert Rocky Intertidal Zone. American Naturalist - AMER
NATURALIST, 127, March 1986.
[137] J. E. Fauth. Interactive Effects of Predators and Early Larval Dynamics of
the Treefrog Hyla Chrysocelis. Ecology, 71(4), August 1990.
117
[138] W. B. Worthen and J. L. Moore. Higher-Order Interactions and Indirect
Effects: A Resolution Using Laboratory Drosophila Communities. The
American Naturalist, 138(5), November 1991.
[139] T. J. Case and E. A. Bender. Testing for Higher Order Interactions. The
American Naturalist, 118(6), December 1981.
[140] J. E. Losey and R. F. Denno. Positive PredatorâPredator Interactions:
Enhanced Predation Rates and Synergistic Suppression of Aphid
Populations. Ecology, 79(6), 1998.
[141] Z. Zhang et al. Spatial heterogeneity and co-occurrence patterns of human
mucosal associated intestinal microbiota. The ISME journal, 8(4):881–893,
April 2014.
[142] J. Qin et al. A human gut microbial gene catalogue established by
metagenomic sequencing. Nature, 464(7285):59–65, March 2010.
[143] R. M. May. Will a Large Complex System be Stable? Nature,
238(5364):413–414, August 1972.
[144] J. E. Goldford et al. Emergent simplicity in microbial community assembly.
Science, 361(6401):469, August 2018.
[145] R. DâAndrea et al. Emergent neutrality in consumer-resource dynamics.
bioRxiv, pp. 710541, January 2019.
[146] G. T.-W. Shaw et al. MetaMIS: a metagenomic microbial interaction
simulator based on microbial community profiles. BMC Bioinformatics,
17(1):488, November 2016.
[147] M. Westerfield and M. Westernfield. The zebrafish book : a guide for the
laboratory use of zebrafish (Danio rerio). University of Oregon, Eugene, 5th
ed. edition, 2007.
[148] W. Z. Stephens et al. The composition of the zebrafish intestinal microbial
community varies across development. The ISME journal, 10(3):644–654,
2016.
[149] C. D. Nadell et al. Spatial structure, cooperation and competition in biofilms.
Nature Reviews Microbiology, 14(9), September 2016.
[150] E. Trudnowska et al. Marine snow morphology illuminates the evolution of
phytoplankton blooms and determines their subsequent vertical export.
Nature Communications, 12(1), May 2021.
118
[151] C. M. Flintrop et al. Embedding and slicing of intact in situ collected marine
snow. Limnology and Oceanography: Methods, 16(6):339–355, 2018.
[152] C. M. Prieto-Barajas et al. Microbial mat ecosystems: Structure types,
functional diversity, and biotechnological application. Electronic Journal of
Biotechnology, 31:48–56, January 2018.
[153] R. Durrett and S. Levin. Spatial Aspects of Interspecific Competition.
Theoretical Population Biology, 53(1):30–43, February 1998.
[154] P. Amarasekare. Competitive coexistence in spatially structured
environments: a synthesis. Ecology Letters, 6(12), 2003.
[155] B. Kerr et al. Local dispersal promotes biodiversity in a real-life game of
rockâpaperâscissors. Nature, 418(6894), July 2002.
[156] B. Shorrocks and M. Bingley. Priority effects and species coexistence:
experiments with fungal-breeding Drosophila. 1994.
[157] W. P. Sousa. Experimental Investigations of Disturbance and Ecological
Succession in a Rocky Intertidal Algal Community. Ecological Monographs,
49(3), 1979.
[158] E. K. Costello et al. The Application of Ecological Theory Toward an
Understanding of the Human Microbiome. Science, 336(6086), June 2012.
[159] D. Sundarraman et al. Higher-Order Interactions Dampen Pairwise
Competition in the Zebrafish Gut Microbiome. mBio, 11(5):e01667–20,
October 2020.
[160] J.-F. Sicard et al. Interactions of Intestinal Bacteria with Components of the
Intestinal Mucus. Frontiers in Cellular and Infection Microbiology, 7, 2017.
[161] C. Robbe et al. Structural diversity and specific distribution of O-glycans in
normal human mucins along the intestinal tract. Biochemical Journal, 384(Pt
2):307–316, December 2004.
[162] J. M. Rhodes. Colonic mucus and mucosal glycoproteins: the key to colitis
and cancer? Gut, 30(12), December 1989.
[163] T. J. Smith et al. A mucin-regulated adhesin determines the intestinal
biogeography and inflammatory character of a bacterial symbiont. bioRxiv.
[164] M. A. McGuckin et al. Mucin dynamics and enteric pathogens. Nature
Reviews. Microbiology, 9(4):265–278, April 2011.
[165] G. C. Hansson. Role of mucus layers in gut infection and inflammation.
Current Opinion in Microbiology, 15(1):57–62, February 2012.
119
[166] J. M. H. Larsson et al. A complex, but uniform O-glycosylation of the human
MUC2 mucin from colonic biopsies analyzed by nanoLC/MSn. Glycobiology,
19(7):756–766, July 2009.
[167] N. Juge. Microbial adhesins to gastrointestinal mucus. Trends in
Microbiology, 20(1):30–39, January 2012.
[168] K. E. Carothers et al. The Streptococcal Protease SpeB Antagonizes the
Biofilms of the Human Pathogen Staphylococcus aureus USA300 through
Cleavage of the Staphylococcal SdrC Protein. Journal of Bacteriology,
202(11).
[169] R. Randal Bollinger et al. Human secretory immunoglobulin A may
contribute to biofilm formation in the gut. Immunology, 109(4), 2003.
[170] S. Mukherjee and B. L. Bassler. Bacterial quorum sensing in complex and
dynamically changing environments. Nature Reviews. Microbiology,
17(6):371–382, June 2019.
[171] S. Elias and E. Banin. Multi-species biofilms: living with friendly neighbors.
FEMS Microbiology Reviews, 36(5):990–1004, September 2012.
[172] A. H. Rickard et al. Autoinducer 2: a concentration-dependent signal for
mutualistic bacterial biofilm growth. Molecular Microbiology, 60(6), 2006.
[173] T. D. Lawley and A. W. Walker. Intestinal colonization resistance.
Immunology, 138(1):1–11, January 2013.
[174] V. Thomas et al. Molecular Characterization and Spatial Analysis of a
Simplified Gut Microbiota Displaying Colonization Resistance against
Clostridium difficile. Microbial Ecology in Health and Disease, 14(4), January
2002.
[175] H. Knecht et al. Effects of β-Lactam Antibiotics and Fluoroquinolones on
Human Gut Microbiota in Relation to Clostridium difficile Associated
Diarrhea. PLOS ONE, 9(2), February 2014.
[176] J. Y. Yoo et al. Gut Microbiota and Immune System Interactions.
Microorganisms, 8(10):1587, 2020.
[177] D. Zheng et al. Interaction between microbiota and immunity in health and
disease. Cell Research, 30(6), June 2020.
[178] E. M. Brown et al. The role of the immune system in governing host-microbe
interactions in the intestine. Nature Immunology, 14(7), July 2013.
120
[179] S. H. Lam et al. Development and maturation of the immune system in
zebrafish, Danio rerio: a gene expression profiling, in situ hybridization and
immunological study. Developmental and Comparative Immunology,
28(1):9–28, January 2004.
[180] B. Ruder et al. Tumour Necrosis Factor Alpha in Intestinal Homeostasis and
Gut Related Diseases. International Journal of Molecular Sciences,
20(8):E1887, April 2019.
[181] L. Marjoram et al. Epigenetic control of intestinal barrier function and
inflammation in zebrafish. Proceedings of the National Academy of Sciences
of the United States of America, 112(9):2770–2775, March 2015.
[182] F. Ellett et al. mpeg1 promoter transgenes direct macrophage-lineage
expression in zebrafish. Blood, 117(4):e49–56, January 2011.
[183] K. K. Takaki et al. Schistosoma mansoni Eggs Modulate the Timing of
Granuloma Formation to Promote Transmission. Cell Host & Microbe,
29(1):58–67.e5, January 2021.
[184] L. Ramakrishnan. Looking Within the Zebrafish to Understand the
Tuberculous Granuloma. In M. Divangahi, editor, The New Paradigm of
Immunity to Tuberculosis, Advances in Experimental Medicine and Biology,
pp. 251–266. Springer, New York, NY, 2013.
[185] H. M. Isles et al. Pioneer neutrophils release chromatin within in vivo swarms.
eLife, 10:e68755, July 2021.
[186] L. Cantas et al. Impact of antibiotic treatments on the expression of the R
plasmid tra genes and on the host innate immune activity during pRAS1
bearing Aeromonas hydrophila infection in zebrafish (Danio rerio). BMC
Microbiology, 12:37, March 2012.
121